深蓝背后

Behind Deep Blue


 

 


深蓝背后

Behind Deep Blue


打造击败世界象棋冠军计算机_ _ _ _ _ _ _

BUILDING THE COMPUTER THAT DEFEATED THE WORLD CHESS CHAMPION

-徐雄旭_ _

FENG-HSIUNG HSU

乔恩·克莱因伯格 (Jon Kleinberg) 的新前言和作者的新序言

With a new foreword by Jon Kleinberg and a new preface by the author

普林斯顿大学出版社

Princeton University Press

普林斯顿大学和牛津大学

Princeton and Oxford

 

 

版权所有© 2002 许凤雄

Copyright © 2002 by Feng-hsiung Hsu

新平装本版本序言,版权所有 © 2022 普林斯顿大学出版社 新平装版本序言,版权所有 © 2022 徐凤雄

Foreword to the new paperback edition, copyright © 2022 by Princeton University Press and preface to the new paperback edition, copyright © 2022 by Feng-hsiung Hsu

由普林斯顿大学出版社出版,地址

:41 William Street, Princeton, New Jersey 08540

Published by Princeton University Press,

41 William Street, Princeton, New Jersey 08540

英国:普林斯顿大学出版社,

99 Banbury Road, Oxford OX2 6JX

In the United Kingdom: Princeton University Press,

99 Banbury Road, Oxford OX2 6JX

版权所有

All Rights Reserved

2002年首次由普林斯顿大学出版社出版

First published by Princeton University Press in 2002

第一本平装本印刷,附新后记,2004 年

First paperback printing, with a new afterword, 2004

新平装本,附有 Jon Kleinberg 的新前言和作者的新序言,2022 年

New paperback edition, with a new foreword by Jon Kleinberg and a new preface by the author, 2022

论文 ISBN 9780691235134

Paper ISBN 9780691235134

美国国会图书馆控制号:2021948921

Library of Congress Control Number: 2021948921

ISBN(电子书)9780691235141

ISBN (ebook) 9780691235141

1.0版本

Version 1.0

大英图书馆出版编目数据可用

British Library Cataloging-in-Publication Data is available

封面艺术:Sukutangan

Cover art by Sukutangan

普林斯顿大学出版社

press.princeton.edu

 

 

内容

Contents


  1. 乔恩·克莱因伯格的前言
  2. Foreword by Jon Kleinberg
  3. 2022 年版序言
  4. Preface to the 2022 Edition
  5. 第一版序言
  6. Preface to the First Edition
  7. 致谢
  8. Acknowledgments
  9. 国际象棋记谱法
  10. Chess Notation
  11. 序幕
  12. Prologue
  13. 第1     表演时间!
  14. CHAPTER 1     Show Time!
  15. 卡内基梅隆大学
  16. Carnegie Mellon
  17. 第二     麻烦制造者的办公室
  18. CHAPTER 2     An Office of Troublemakers
  19. 第3     冒险
  20. CHAPTER 3     Taking the Plunge
  21. 第4     并非如此的国际象棋机器
  22. CHAPTER 4     The Chess Machine That Wasn’t
  23. 第5     第一机器宗师之争
  24. CHAPTER 5     The Race for First Machine Grandmaster
  25. 第6     “敲门,敲门。谁在那儿?”
  26. CHAPTER 6     “Knock, Knock. Who’s There?”
  27. 间奏曲
  28. Intermezzo
  29. 第7     第一次约会历史
  30. CHAPTER 7     First Date with History
  31. 国际商业机器公司
  32. IBM
  33. 第8     我们需要一个新名字
  34. CHAPTER 8     We Need a New Name
  35. 第9     抚养婴儿
  36. CHAPTER 9     Bringing Up the Baby
  37. 第10   活生生的珠穆朗玛峰
  38. CHAPTER 10   A Living Mount Everest
  39. 11   重组
  40. CHAPTER 11   Retooling
  41. 第十二圣杯_   
  42. CHAPTER 12   The Holy Grail
  43. 结语
  44. Epilogue
  45. 第13   国际象棋之后的生活
  46. CHAPTER 13   Life after Chess
  47. 2004年平装本后记
  48. Afterword to the 2004 Paperback Edition
  49. 附录
  50. Appendixes
  51. 附录A A   来自台湾的小伙子
  52. APPENDIX A   A Lad from Taiwan
  53. A附录B   部分比赛得分
  54. APPENDIX B   Selected Game Scores
  55. 附录C   延伸阅读
  56. APPENDIX C   Further Reading

 

 

前言

Foreword


人工智能有很多隐喻,首先是“人工智能”一词本身。在这些比喻中,最令人回味的比喻之一始终是与外星智能的类比——真正与强大的人工智能互动就是与感觉超凡脱俗的东西接触。

There are many metaphors for artificial intelligence, beginning with the term “artificial intelligence” itself. And among these metaphors, one of the most evocative has always been the analogy with alien intelligence—that to truly interact with powerful artificial intelligence is to make contact with something that feels otherworldly.

科幻小说为我们提供了心理图像,让我们了解在这样的遭遇中会发生什么。外星飞船坐在我们面前,不透明,冷漠,等待着。由于缺乏任何共同的经验,我们的沟通必须结构化和风格化;我们发明对话规则。我们向它闪烁一系列灯光:2、3、5。灯光又闪烁回来:7、11、13。控制室舞台外的人类科学家从终端跳回来,惊呼起来。它知道素数。

Science fiction has provided us with mental images for what to expect in such an encounter. The alien spacecraft sits in front of us, opaque, impassive, waiting. Lacking any shared experience with it, our communication has to be structured and stylized; we invent rules for conversation. We flash a series of lights at it: 2, 3, 5. Lights blink back in return: 7, 11, 13. The human scientists offstage in the control room jump back from their terminals, exclaiming. It knows about the prime numbers.

在这些故事的陪伴下长大的我,从来没有意识到它们与 1996 年和 1997 年观看国际象棋世界冠军加里·卡斯帕罗夫 (Garry Kasparov) 与深蓝 (Deep Blue) 比赛的经历有多么接近。两个强大的智能体,彼此深感难以理解,坐下来通过神秘的世界进行交流。 ,国际象棋走法的程式化语言。卡斯帕罗夫在棋盘上下了一步棋,经过几分钟和几十亿次计算后,一步棋得到了回应。这两个实体正在相互交谈,每个人都试图了解对方。我们无法理解,而且在某种非常真实的意义上无法理解,深蓝到底是如何做出它的行动选择的。但我们确实知道,无论它如何下棋,它都与桌子对面的人类对手下棋的方式完全不同。这样一来,就很难摆脱外星人存在的感觉。悄悄地进入比赛场地——控制室台下评论员莫里斯·阿什利(Maurice Ashley)和亚西尔·塞拉万(Yasser Seirawan)经常发出的感叹强化了这种感觉:它知道位置典当牺牲;它知道如何向第二侧发起进攻;它知道阻止对手反击的重要性。这些都是我们认为是为了讨论国际象棋策略而创建的人类概念,我们看到它们是从深蓝游戏的计算中浮现出来的。

Growing up with these stories, I never appreciated how closely they would match the experience of watching the world chess champion Garry Kasparov play against Deep Blue in 1996 and 1997. Two powerful intelligences, profoundly incomprehensible to each other, sit down and communicate through the cryptic, stylized language of chess moves. Kasparov puts a move on the board, and after a few minutes and a few billion computations, a move comes back in reply. The two entities are conversing with each other, each seeking to understand the other. We do not fathom, and in some very real sense cannot fathom, exactly how Deep Blue has arrived at its choice of moves. But we do know that however it is managing to play chess, it is utterly different from how its human opponent across the table is playing chess. And in this way, it is hard to escape the feeling that an alien presence has stealthily entered the playing venue—a feeling reinforced by regular exclamations from commentators Maurice Ashley and Yasser Seirawan offstage in the control room: it knows about positional pawn sacrifices; it knows how to open an attack on a second flank; it knows the importance of blocking its opponent’s counterplay. These are all human concepts that we thought we had created to talk about chess strategy, and we were seeing them emerge from the computations that went into Deep Blue’s play.

《深蓝》故事的微妙之处在于,这种与外星智慧生物第一次接触的比喻部分是恰当的,部分是一种幻觉。这在一定程度上是恰当的,因为深蓝得出其走法的过程与人类在国际象棋游戏中思考走法的经验非常不同。但这在一定程度上也是一种幻想,因为它忽略了构建深蓝的人类团队的作用,以及开发计算机国际象棋所基于的算法的人类研究领域。我们无法对机器如何选择任何特定的动作给出完整的解释,但我们对它的搜索机制以及驱动其决策的想法和启发法了解很多;我们创造了这些东西。

Part of the subtlety to the story of Deep Blue is the way in which this metaphor of first contact with an alien intelligence is partly apt and partly an illusion. It is partly apt because the process by which Deep Blue arrived at its moves was so different from the human experience of thinking about moves in a chess game. But it is also partly an illusion, because it omits the role of the human team that built Deep Blue, and the human fields of study that developed the algorithms on which computer chess is based. We cannot give a full explanation for how the machine chose any specific move, but we know a lot about the mechanisms by which it searches and the ideas and heuristics that drive its decisions; we created these things.

强大人工智能的人类和外星人维度之间的这种紧张关系在徐凤雄讲述他和他的同事创建深蓝的多年旅程中发挥着关键作用。自从卡斯帕罗夫坐在机器对面的棋盘上二十五年以来,我们已经看到这种紧张关系在许多采用人工智能的领域中体现出来——在搜索、推荐和图像识别等在线应用程序中;在人工智能应用领域中。在医学、教育、招聘、贷款、政治进程和法律体系等高风险社会环境中。虽然这些表现形式多种多样,但有必要强调三个反复出现的主题,特别是我们在深蓝的发展中看到的主题:规模的数量增加所产生的质变;解释人工智能应用程序做出的决策的挑战;以及人类参与者在更大的人工智能生态系统中的持续作用。

This tension between the human and alien dimensions of powerful AI plays a key role in Feng-hsiung Hsu’s narrative of the multiyear journey that he and his colleagues followed to create Deep Blue. And in the twenty-five years since Kasparov sat across the chessboard from the machine, we have seen this tension manifest itself across many of the domains where AI is adopted—in online applications like search, recommendation, and image recognition; and in high-stakes societal contexts like medicine, education, hiring, lending, the political process, and the legal system. While these manifestations are diverse, it is useful to highlight three recurring themes in particular that we saw foreshadowed by the development of Deep Blue: the qualitative changes produced by quantitative increases in scale; the challenge of explaining the decisions made by AI applications; and the continuing role of human participants in the larger AI ecosystem.

第一个主题在卡斯帕罗夫在第一场深蓝比赛中说出的口头禅中得到了很好的概括:“数量变成质量”。当然,这从来都不是纯粹数量的问题,因为人类对人工智能算法和架构的定性洞察对于他们的成功至关重要。但是,我们如何从逐步理解的算法开始,在概念上仍然存在一些难以捉摸的地方,当我们将数十亿个这样的步骤串在一起时,我们会看到很难用以下术语来解释的紧急行为:步骤本身。正如 Hsu 在讲述这个故事时所写的那样,“有时,每秒搜索 2 亿个位置会产生不寻常的新解决方案。” 近年来,这种效应变得更加强大,无论是在国际象棋领域,还是在部署人工智能的许多其他领域,当前的国际象棋引擎(例如 DeepMind 的 AlphaZero)通过直接从游戏规则开始进行自我训练,已经达到了巨大的强度。 。

The first of these themes is well-summarized in the mantra uttered by Kasparov in the first Deep Blue match, that “quantity becomes quality.” Of course, it has never been a matter of pure quantity, since the qualitative insights that humans put into the algorithms and the architectures of AI are crucial for their success. But there remains something conceptually elusive about the way we can start with an algorithm that we understand at a step-by-step level, and when we string together billions of these steps, we see emergent behaviors that are very hard to explain in terms of the steps themselves. As Hsu writes in telling the story, “Sometimes, searching two-hundred million positions per second produces unusual new solutions.” This effect has become even more potent in recent years, both in chess, where current chess engines like DeepMind’s AlphaZero have achieved enormous levels of strength by training against themselves starting directly from the rules of the game, and in many other areas where AI is deployed.

数量可以转化为质量这一事实也是我们从物理世界中知道的一个原理。想想螺旋桨。当它关闭时,您可以用手移动钝的金属刀片并查看其旋转。但从手工模拟来看,很难理解如何大幅提高同一个螺旋桨的速度,让它能够完成看似神奇的事情,将重型飞机升离地面。或者当它这样做时会变得多么危险。或者事实是,在这样的速度下,人眼基本上看不见它。

The fact that quantity can become quality is a principle we know from the physical world as well. Think of a propeller. When it’s turned off, you can move the blunt metal blade by hand and see its rotations. But from the hand-simulation, it’s hard to understand how dramatically increasing this same propeller’s speed can allow it to do things that seem magical, lifting heavy aircraft off the ground. Or how dangerous it becomes when it’s doing that. Or the fact that at these speeds it will be essentially invisible to the human eye.

当我们尝试手动模拟为现代人工智能应用提供动力的算法并推理它们可能做什么时,我们也会遭遇类似的想象力失败。随着他们的力量不断增强,他们开始理解我们的语言,解释我们的文字,解释他们在照片中看到的内容,自主引导我们的车辆,诊断我们的医疗状况,并发现新的数学事实和科学现象。他们还会在我们认为相对没有风险的情况下犯下破坏性错误。我们很难预测他们的成功和失败,以及当我们对他们使用的方法或他们训练的数据进行微小改变时,他们会如何变化。所有这一切都发生在通常主要在后台运行的系统中,在搜索结果页面不起眼的白屏或手机语音识别系统的时尚界面后面。这样一来,AI也像螺旋桨一样变得神奇、危险、隐形。

We suffer similar failures of imagination when we try to hand-simulate the algorithms that power modern AI applications and to reason about what they might do. As they have grown in strength, they have come to understand our speech, paraphrase our writing, explain what they see in our photos, autonomously guide our vehicles, diagnose our medical conditions, and discover new mathematical facts and scientific phenomena. They also make damaging errors in situations that we would have imagined as relatively free of risk. We struggle to anticipate both their successes and their failures, and how they vary as we make small changes to the methods they use or the data they are trained on. And all this is happening with systems that are often operating largely in the background, behind the unassuming white screen of a search results page or the sleek interface of a phone’s speech recognition system. In this way, like the propeller, AI has also become magical, dangerous, and invisible.

理解复杂人工智能系统如何产生不同行为的困难支撑了第二个中心主题,即可解释性的挑战:为人类可以理解的系统行为提供解释。经常这是构建这些应用程序的工程师迫切需要的功能,他们面临着在像深蓝这样复杂的系统中查找和纠正错误的挑战。但有时,可解释性问题也会影响到系统的最终用户,例如,当我们想知道为什么算法系统会给出特定的医疗诊断或贷款申请的特定决定时。

The difficulty in understanding how different behaviors emerge from a complex AI system underpins a second central theme, the challenge of interpretability: to provide explanations of a system’s behavior that are comprehensible to humans. Often this is a functionality urgently needed by the engineers who build these applications, faced with the challenge of finding and correcting bugs in a system as complex as Deep Blue. But sometimes the problem of interpretability reaches the end user of the system as well—for example, when we want to know why an algorithmic system gave us a particular medical diagnosis, or a particular decision on a loan application.

这种可解释性的崩溃在卡斯帕罗夫和深蓝比赛的核心情节之一中发挥了关键作用,这场比赛围绕计算机做出的一个动作引发了一场争议和辩论:计算机在第二场比赛的第二局中的决定它的主教是方格e4。传统上,计算机偏爱在其他条件相同的情况下赢得对手棋子和棋子的走法,但深蓝在这一步棋中拒绝了这种诱惑,而是用一种安静但致命的防御走法扼杀了卡斯帕罗夫潜在反击的希望。此举违反了当时国际象棋程序应如何运作的普遍思维模式,并且使卡斯帕罗夫团队的计划失去了平衡。那么深蓝是如何做出这一举动的决定的呢?正如许在他的比赛故事中所总结的那样,在某些基本层面上,这个问题没有简单的答案。在考虑这一举措的几分钟内,深蓝执行了数十亿次计算。即使有完整的推理记录,也需要比人一生更长的时间才能读完。最后,对于计算为何得出这样的结论,可能仍然没有简短的解释。在某些情况下,即使专家拥有所有的知识和动机,计算机的决策也可能对人类来说本质上是难以理解的。

This breakdown of interpretability plays a key role in one of the central episodes of the matches between Kasparov and Deep Blue, a hurricane of controversy and debate around a single move made by the computer: its decision in the second game of the second match to play its bishop to the square e4. Computers had traditionally favored moves that—all else being equal—won their opponents’ pieces and pawns, but on this move Deep Blue spurned such temptations and instead smothered Kasparov’s hopes of a potential counterattack with a quiet but deadly defensive move. The move violated the ubiquitous mental model for how chess programs at the time were supposed to work, and it threw the planning of Kasparov’s team off balance. So how did Deep Blue arrive at the decision to make this move? At some fundamental level, as Hsu concludes in his story of the match, there is no easy answer to this question. In the few minutes while it was considering this move, Deep Blue performed billions of computations. Even with a complete transcript of its reasoning, it would take more than a human lifetime to read through it; and at the end, there would likely still be no short explanation for why the computation reached the conclusion it did. In some situations, even with all the knowledge and motivation that experts can bring to bear, a computer’s decision may be inherently incomprehensible to human beings.

因此,缺乏可解释性有时是不可避免的,但这并不是全有或全无的情况。我们可以努力创建能够增加解释程度的结构,从而为系统建立可靠性和信任。深蓝的故事也是这个过程的故事,并且在人工智能应用的各个地方仍在继续。因为如果没有一些信任系统的基础,我们就会对所看到的决策失去信心;我们对自己的行动方针产生怀疑。冲突转向内心,我们开始打败自己。

So a lack of interpretability can sometimes be inevitable, but it is not all-or-nothing; we can work to create structures that increase the extent of explanation, that build dependability and trust into the system. The story of Deep Blue is the story of this process as well, and it is still ongoing throughout the places where AI is used. Because without some basis for trusting the system, we lose faith in the decisions we see; we form doubts about our own courses of action. The conflict turns inward, and we start to defeat ourselves.

这给我们带来了第三个主题——人工智能的人性维度。需要在其预期的背景下设计人工智能应用程序,以在每个领域实现可靠性和信任,这表明至少目前,其使用将继续具有强烈的人性化成分。它让我们欣赏到人类仍然以多种方式陪伴着我们与人工智能的互动——人工智能何时被构建、何时被使用、何时被解释——以及这些角色的重要性。

And this brings us to the third theme—the human dimension of AI. The need to design AI in the context of its intended applications, to achieve dependability and trust in each domain, suggests that at least for now there will continue to be a strongly human component to its use. It makes us appreciate the myriad ways in which humans still chaperone our interactions with AI—when they are built, when they are used, and when their results are interpreted—and the importance these roles can have.

我们可以预期,专家协调员、最终用户和底层人工智能系统之间的三角交互会越来越多样化——例如,医护人员、患者和医疗人工智能应用程序之间;或者一个老师、一个学生和一件教育技术。我对这种由三方陪同的人机交互动态的最早体验之一也涉及深蓝。1995 年夏天,作为一名研究生,我感到有点追星,在 IBM Watson 暑期实习生演示的一个下午坐下来玩了 Deep Blue 原型。除了游戏本身之外,我还记得在棋盘另一边与默里·坎贝尔(Murray Campbell)进行的对话,当时他正在为计算机移动棋子,在不同的点上讨论深蓝所看到和期望的内容,并在不可避免的结局中以最温和的方式进行解释深蓝的国王可能会从方格g8一直到a7上的藏身之处,而我不会通过我正在寻找的永久检查来获得平局。

We can likely expect an increasing variety of these triangular interactions between an expert facilitator, an end user, and an underlying AI system—for example, between a health-care worker, a patient, and a medical AI application; or a teacher, a student, and a piece of educational technology. One of my earliest experiences with this chaperoned three-party dynamic of human-AI interaction involved Deep Blue as well. As a graduate student in summer 1995, feeling slightly starstruck, I sat down to play the Deep Blue prototype during an afternoon of demos for summer interns at IBM Watson. Beyond the game itself, I remember the conversation with Murray Campbell on the other side of the board as he moved the pieces for the computer, discussing at various points what Deep Blue was seeing and expecting, and at the inevitable end, explaining in the gentlest possible way that Deep Blue’s king was going to make it all the way across the board from the square g8 to a hiding place on a7, and I wasn’t going to get the draw by perpetual check that I was looking for.

增加人类人工智能生态系统的丰富性,使人工智能更安全、更值得信赖、使用时破坏性更小,这一目标将继续需要我们所有的创新。在交互中添加适当的人类背景层次是实现这一目标的一个重要方向。另一个潜在的方向是创建行为更加人性化的人工智能系统,这样直接与它们合作的体验可能会开始感觉不那么陌生。在国际象棋方面,我和几位同事(阿什顿·安德森 (Ashton Anderson)、里德·麦克罗伊-杨 (Reid Mcllroy-Young) 和悉达多·森 (Siddhartha Sen))通过创建一个新的国际象棋程序 Maia 探索了这个想法,该程序可以进行调整以匹配人类棋手的动作。比标准国际象棋引擎能够更准确地识别不同的技能水平。这是另一种优化的例子——不是为了在解决任务时更接近完美,而是为了更接近人类处理任务的方法。

Increasing the richness of the human-AI ecosystem in a way that makes AI safer, more trustworthy, and less disruptive to use is a goal that will continue to require all our innovation. Adding appropriate layers of human context to the interaction is one important direction toward this goal. Another potential direction is the creation of AI systems that are more humanlike in their behavior, so that the experience of working with them directly may begin to feel less alien. In the context of chess, several colleagues (Ashton Anderson, Reid Mcllroy-Young, and Siddhartha Sen) and I have explored this idea by creating a new chess-playing program, Maia, that can be tuned to match the moves of human players of different skill levels much more accurately than standard chess engines can achieve. It is an example of a different kind of optimization—not aimed at coming closer to perfection in solving the task but at more closely resembling the human approach to the task.

在这一点上,正如在许多相关问题中一样,国际象棋是一个有效的模型系统,正是因为它是如此结构化、如此明确定义——这个独立的世界足够复杂,足以迫使我们面对和解决严峻的挑战,但又足够清晰,足以为我们何时成功提供相对具体的感觉。深蓝可以说是这一研究范式历史上最引人注目的一章,它预示了更广泛的人工智能领域仍在积极解决的一系列挑战:当我们引入这些系统时,如何预测它们的后果进入一系列风险越来越高的领域;如何确保我们能够向其生活受到影响的个人解释他们的决定;以及如何确保人工智能的行为与人类的行为之间富有成效的共存。

In this, as in so many related questions, chess is a potent model system precisely because it is so structured, so well-defined—a self-contained world that is complex enough to force us to confront and solve hard challenges but clear-cut enough to provide a relatively concrete sense for when we have succeeded. Deep Blue is arguably the most visible chapter in the history of this research paradigm, and it foreshadowed a set of challenges that the broader field of AI is still actively in the middle of resolving: how to anticipate the consequences of these systems as we introduce them into a series of increasingly high-stakes domains; how to make sure we can explain their decisions to the individuals whose lives they affect; and how to ensure productive coexistence between the behavior of AI and the behavior of human beings.

在可预见的未来,这些问题将继续挑战我们。不过,现在让我们回想一下计算历史上的一个时刻,当时两个智能生物——一个是人类及其支持团队;另一个是人类和他的支持团队。一台计算机和它的设计者——在棋盘上短暂交锋,在分道扬镳之前努力相互理解。

These questions will continue to challenge us for the foreseeable future. For now, though, let’s think back to a moment in the history of computing when two intelligent beings—a human with his support team; a computer with its designers—briefly engaged across a chessboard and struggled to understand each other before going their separate ways.

——乔恩·克莱因伯格

—Jon Kleinberg

 

 

2022 年平装本序言

Preface to the 2022 Paperback


截至 2021 年撰写本文时,距离深蓝在 1997 年战胜国际象棋世界冠军加里·卡斯帕罗夫 (Garry Kasparov) 的比赛已经过去了 24 年。摩尔定律,即每个集成电路的元件数量每两年翻一番,其用处已接近尾声。我个人认为人类与计算机的国际象棋比赛是人类扮演表演者角色和人类扮演工具制造者角色之间的竞争。二十四年来,人类作为表演者和人类作为工具制造者之间的竞争在更多的游戏中向工具制造者倾斜。摩尔定律增强的计算能力发挥了一定作用,但工具制造商也推出了更多更好的工具。

At the time of this writing in 2021, twenty-four years have passed since Deep Blue won the match with Garry Kasparov, the World Chess Champion, in 1997. Moore’s Law, that the number of components per integrated circuit doubles every two years, is near the end of its usefulness. I personally view human vs. computer chess competitions as competitions between human in the role of performer and human in the role of toolmaker. In twenty-four years, the competitions between human as a performer and human as a toolmaker have skewed in favor of the toolmaker in many more games. The enhanced computation power from Moore’s Law plays a part, but the toolmaker has also come up with more and better tools.

当深蓝在 1997 年击败加里·卡斯帕罗夫时,它使用 480 个定制的国际象棋芯片每秒搜索大约 2 亿个棋局,每个芯片每秒搜索多达 300 万个棋局。如果我们将一层定义为半步国际象棋棋步,或者两个国际象棋棋手之一下棋,那么“深蓝”平均至少达到 12 层深度,沿强迫线达到 40 多个层。深蓝定制国际象棋芯片所用技术的最小线宽比当前技术水平粗约 70 倍(350 nm 与 5 nm)。目前线宽的减小可以提高面密度并提高晶体管速度。今天重新设计和优化的单芯片国际象棋计算机很可能比深蓝快 1000 倍。这样的单芯片国际象棋计算机应该搜索至少 16 层深度,并沿着强制线搜索超过 50-60 层。

When Deep Blue defeated Garry Kasparov in 1997, it searched about 200 million chess positions per second, using 480 custom chess chips, with each chip searching up to 3 million positions per second. If we define a ply as a half chess move, or one of the two chess players playing a move, Deep Blue reached on average at least 12 plies deep and more than 40 plies along forcing lines. The minimum line width in the technology used for Deep Blue’s custom chess chip was about 70 times (350 nm vs 5 nm) coarser than the current state of the art. The line width reduction available today leads to both areal density improvement and faster transistors. A single chip chess computer, redesigned and optimized today, could very well be 1,000 times faster than Deep Blue. Such a single chip chess computer should search at least 16 plies deep and more than 50–60 plies along forcing lines.

在计算机国际象棋历史的早期,许多研究人员认为计算机有必要选择性地修剪掉毫无希望的举动,以便它可以沿着“有趣”的路线进行足够深入的搜索,以与人类竞争。这是选择性修剪,早期通常是通过临时手写规则来完成的。从 20 世纪 70 年代末开始,人们逐渐认识到,头脑简单的强力搜索(搜索每一个合法的动作,并接受正常的alpha-beta 剪枝)比早期的选择性剪枝程序更容易编程,而且实际上表现得更好。搜索深度足以与人类竞争的问题仍然存在。一种可能的替代方案是通过更深入地搜索“有趣”的行来花费更少的时间来搜索没有希望的动作,而不是像选择性修剪那样完全修剪掉没有希望的动作。选择性扩展,结合了搜索强制移动(例如,检查逃避或夺回移动)的思想(例如更深一层),作为一种风险较小的方式被引入,以允许程序沿着强制线进行更深入的搜索。深蓝很大程度上遵循了这种范式,尽管有更高级的选择性扩展形式。特别是,引入了测试搜索(常规搜索不一定需要的搜索)来衡量动作的力度。例如,如果通常在深度大大降低的情况下进行的测试搜索表明只有一个“好”棋步是可能的,那么该“好”棋步就值得进行更深入的搜索。这构成了奇异扩展背后的基本思想,我在 1986 年观察到另外两个国际象棋程序盲目地进入一长串单向强制移动序列而没有任何一个程序实现最终结果后引入了该思想。我必须承认我的懒惰是采用测试搜索的主要原因因为几乎不需要额外的编程。

Early in computer chess history, many researchers thought it was necessary for the computer to selectively prune away unpromising moves so that it could search deep enough along “interesting” lines to compete with humans. This was selective pruning, usually done with ad hoc handwritten rules in the early days. From the late 1970s onward, it became clear that simple-minded brute force search, which searches every single legal move, subject to normal alpha-beta pruning, was easier to program and actually performed better than the early selective pruning programs. The problem of searching deep enough to compete with humans still exists. Instead of pruning away unpromising moves completely as in selective pruning, a possible alternative is to spend less time searching the unpromising moves by just searching “interesting” lines deeper. Selective extensions, incorporating the idea of searching forced moves (for example, check evasion or recapture moves) say one ply deeper, were introduced as a less risky way to allow the program to search deeper along forcing lines. Deep Blue largely followed this paradigm, albeit with more advanced forms of selective extensions. In particular, test searches, searches that are not necessarily needed for regular search, were introduced to measure the forcefulness of the moves. For example, if test searches, typically done at much reduced depth, suggest only one single “good” move is possible, then the “good” move deserves to be searched deeper. This forms the basic idea behind singular extensions, which I introduced in 1986 after observing two other chess programs blindly going into a long one-way sequence of forced moves without either program realizing the eventual outcome. I have to admit that my laziness was the main reason behind the adoption of test searches, as very little additional programming was needed.

在 20 世纪 80 年代末和 90 年代初,选择性修剪算法(而不是临时手写规则)通过测试搜索来修剪掉没有希望的动作,被证明是有效且相对稳健的。与移动的常规搜索相比,对修剪移动的测试搜索的计算成本必须要低得多,这可以通过以大大减少的深度(例如少两层)进行测试搜索来完成。一个例子是空棋修剪,其中测试搜索让一名玩家连续进行两步棋,而对手则在两者之间进行零棋。如果在玩家连续进行两次移动之后,测试搜索对于深度减小的玩家来说“不令人满意”,那么玩家的第一次移动是没有希望的并且可以被修剪掉。如果我们允许测试搜索本身是空移动修剪,我们有所谓的递归空移动修剪。特别是,当搜索深度足够高时,递归空移动修剪可以使起始强力程序的搜索深度加倍。深蓝没有使用递归空步修剪,因为我们认为可能丢失关键步的风险,无论多么轻微,都不是我们想要承担的。

In the late 1980s and early 1990s, selective pruning algorithms, which use, instead of ad hoc handwritten rules, test searches to prune away unpromising moves, proved to be effective and relatively robust. The test searches for pruned moves have to cost way less to compute than the moves’ regular searches, and this can be done by doing the test searches at much reduced depth, say two plies less. One example is null move pruning, where the test searches have one player making two moves in a row, with the opponent thus making a null move in between. If the test search, after a player making two moves in a row, is “unsatisfactory” for the player at reduced depth, then the first move by the player is unpromising and can be pruned away. If we allow test searches themselves to be null move pruned, we have what is known as recursive null move pruning. Recursive null move pruning, in particular, could double the search depth of the starting brute force program when the search depth is sufficiently high. Deep Blue did not use recursive null move pruning, as we deemed the risk of possibly missing key moves, however slight, was not one we wanted to take.

在深蓝战胜卡斯帕罗夫之后,日本将棋和中国围棋被认为是人工智能研究人员的下一个目标。这两种游戏都有更大的分支因素:与国际象棋中的大约三十八步相比,典型的将棋或围棋的走法可能多达数百步。较大的分支因子被认为是类似国际象棋的方法不足以让计算机在这些游戏中击败人类的原因。2007 年,我在IEEE Spectrum 上发表了一篇逆向观点文章,提出计算机国际象棋技术,例如强力搜索和递归空着剪枝,与先进的半导体技术相结合,可以让更难的围棋游戏在十年内被征服。

After Deep Blue’s win over Kasparov, the Japanese game of Shogi and the Chinese game of Go were suggested as the next targets for AI researchers. Both games have much larger branching factors: up to hundreds of moves are possible from a typical Shogi or Go position, compared to the thirty-eight or so moves in chess. The larger branching factors were cited as the reason chess-like approaches would not be sufficient for computers to defeat humans in these games. In 2007, I wrote a contrarian view article in IEEE Spectrum, suggesting that computer chess techniques, such as brute force searching and recursive null move pruning, in combination with advancing semiconductor technology could allow the harder game of Go to be conquered in ten years.

我的日期预测结果是正确的,但路线不同。

My date prediction turned out to be correct, but the route was different.

2017年,AlphaGo在三局比赛中击败了当时世界排名第一的围棋选手柯洁。AlphaGo 在其走法选择中使用了深度神经网络。使用了两个深度神经网络,一个用于评估,另一个用于决定搜索中包含哪一步棋。人类游戏和自我对弈游戏被用来训练深度神经网络。更先进的半导体技术确实在 AlphaGo 的胜利中发挥了作用。在 2016 年与李世石的较早比赛中,《经济学人》报道称 AlphaGo 使用了 1,920 个 CPU 和 280 个 GPU。所以我对日期和技术部分的判断是正确的。三个预测正确两个已经不错了。

In 2017, AlphaGo defeated Ke Jie, the number one ranked Go player in the world at the time, in a three-game match. AlphaGo used deep neural networks in its move selection. Two deep neural networks were used, one for evaluation and the other for deciding which move to include in the search. Human games and self-play games were used to train the deep neural networks. More advanced semiconductor technology did play a part in AlphaGo’s win. In an earlier match against Lee Sedol in 2016, the Economist reported that AlphaGo used 1,920 CPUs and 280 GPUs. So I was right about the date and the technology part. Two predictions correct out of three is not bad.

AlphaGo 中深度神经网络的使用让我想起了深蓝及其前身的棋局评估函数的演变过程。评估函数从我们的第一台国际象棋机 ChipTest 中的简单线性项组合开始,到 Deep Blue 中具有更多项和参数的数学上更复杂的高阶方程。更复杂的评估函数使深蓝能够以不同于其他国际象棋程序的方式看待事物,从而卡斯帕罗夫感到困惑。深度神经网络具有更强的表达能力,使得 AlphaGo 能够改进其评估和走法选择,从而在棋力上比其他程序有显着提高。一些领域的微小改进,通过更大的计算能力放大,会产生巨大的差异。

The use of deep neural networks in AlphaGo reminds me of what happened in the evolution of evaluation function for chess positions in Deep Blue and its predecessors. The evaluation function started from simple linear combinations of terms in ChipTest, our first chess machine, to mathematically more complicated higher order equations with vastly more terms and parameters in Deep Blue. The more complicated evaluation function allowed Deep Blue to see things differently from other chess programs and thus confounded Kasparov. The deep neural networks, with their greater expressive power, allowed AlphaGo to improve its evaluation and its move selection just enough to have a drastic improvement in playing strength over other programs. Small improvements in a few areas, amplified by greater computation power, create huge differences.

 

 

第一版序言

Preface to the First Edition


这本书讲述了我对创建深蓝的冒险的看法,深蓝是第一台在一场严肃的比赛中击败世界象棋冠军的计算机。我于 1985 年开始了这个项目。十二年后,这次冒险结束了,深蓝在人类历史上树立了一个重要的里程碑,并永远改变了我们对如何与计算机相处的看法。

This book recounts my view of the adventure to create Deep Blue, the first computer to defeat the World Chess Champion in a serious match. I started the project in 1985. Twelve years later, the adventure ended with Deep Blue setting a major milestone in human history and forever altering our view of how we would live with the computer.

像深蓝这样的国际象棋机器有着悠久的历史。土耳其人是匈牙利工程师沃尔夫冈·冯·肯佩伦男爵制造的“国际象棋机器”,于 1769 年首次出现在奥地利宫廷。它最著名的受害者可能是拿破仑·波拿巴。然而,土耳其人是一个假象——一个隐藏的人类棋手为它走棋。

Chess machines like Deep Blue have a long history. The Turk, a “chess machine” built by Hungarian engineer Baron Wolfgang von Kempelen, first appeared at the Austrian court in 1769. Its most famous victim was probably Napoleon Bonaparte. The Turk, however, was a fake—a hidden human chess player made the moves for it.

1830 年代,查尔斯·巴贝奇 (Charles Babbage) 梦想建造一台国际象棋机器,为他的分析引擎(第一台机械可编程计算机,但从未完成)吸引资金。相反,他建造了一台井字游戏机。真正的国际象棋机器太难了,尤其是如果你只能用机械部件来建造它的话。

In the 1830s, Charles Babbage dreamed of building a chess machine to attract funding for his Analytic Engine (the first mechanical programmable computer, but never completed). He built a tic-tac-toe machine instead. A real chess machine was too hard, especially if you had to build it with only mechanical parts.

第一台电子可编程计算机 ENIAC 于 1946 年完成。不久之后的 1949 年,克劳德·香农 (Claude Shannon) 发表了一场演讲,为现代国际象棋计算机奠定了基础。当时,许多著名的计算机科学家相信计算机国际象棋问题——创造一台能够击败世界冠军的国际象棋计算机——将在几年内得到解决,但事实证明他们的预测过于乐观。随着时间的推移,在看不到解决方案的情况下,解决计算机国际象棋问题成为计算机科学家的圣杯。到了 20 世纪 80 年代初,也就是香农开创性工作三十多年之后,最好的国际象棋计算机处于美国国家大师水平,距离挑战世界冠军还有很长的路要走。

The first electronic programmable computer, ENIAC, was completed in 1946. Shortly after, in 1949, Claude Shannon gave a lecture that laid the foundation for modern chess computers. At that time, many renowned computer scientists believed that the Computer Chess Problem—creating a chess computer that could beat the World Champion—would be solved within a few years, but their predictions turned out to be overly optimistic. As time went on, and with no solution in sight, solving the Computer Chess Problem became a holy grail for computer scientists. By the early 1980s, over thirty years after Shannon’s seminal work, the best chess computers were playing at US national master level, still a long way from challenging the World Champion.

我接触电脑象棋纯属偶然。1985 年春天,我在卡内基梅隆大学的导师 HT Kung 教授告诉我,我在研究生院的生存存在疑问。几周前,作为一个帮忙,我为我们部门的一位领先的计算机国际象棋专家研究了一个技术问题。他,这位专家,不喜欢我提出的解决方案。我也没有。经过一番思考,我得出的结论是,真正的问题在于他的方法。一种激进的新设计(部分基于竞争对手的想法)要优越得多。在他拒绝了我的新解决方案后,我直言不讳地告诉他,他的方法很差劲,我已经失去了兴趣。孔教授解释说我应该更委婉一些。然后他敦促我写一份技术报告并进行演示来支持我的主张。我之前的一次不检点行为已经惹了够多的麻烦了,差点让我被学校开除……

My involvement with computer chess was by accident. In the spring of 1985, Professor H T Kung, my faculty advisor at Carnegie Mellon, informed me that my survival in graduate school was in doubt. A couple of weeks earlier, as a favor, I looked into a technical problem for a leading computer chess expert in our department. He, the expert, did not like the solution that I came up with. Neither did I. After some thinking, I reached the conclusion that the real issue was with his approach. A radical new design, in part based on his competitors’ ideas, was far superior. After he rejected my new solution, I told him bluntly that his approach was inferior and I had lost interest. Professor Kung explained that I should have been more tactful. He then urged me to write up a technical report and give a presentation to back up my assertion. I already had enough trouble from an earlier indiscretion that almost got me kicked out of the school …

由此开始了创造世界上最强大的国际象棋实体深蓝的奇怪而意想不到的旅程。

Thus began the strange and unexpected journey to create the most powerful chess-playing entity in the world, Deep Blue.

最初,我计划沿着我最喜欢的书《双螺旋:诺贝尔奖得主詹姆斯·D·沃森博士发现 DNA 结构的个人记述》中所走的路线重走这段旅程。然而,当我写作时,我意识到这两本书在范围和性质上会有很大不同。沃森博士的书讲述了二十世纪最伟大的科学发现之一。这本书更多地是关于工程探索而不是科学发现。从本质上讲,工程探索涵盖了更丰富的生活结构。首先必须发现任务背后的技术思想,然​​后将其逻辑完成。这一发现可能来自于运气或灵感,但其余的探索则需要汗水和毅力。因此,这本书的范围更接近查克·耶格尔(Chuck Yeager)的《一马赫的探索》(The Quest for Mach One),他对突破音障的描述。这本书描述了对超越(即使只是短暂的)地球上最好的人类国际象棋棋手的棋艺能力的追求。

Initially, I planned to retrace this journey along the same route as that taken in a favorite book of mine, The Double Helix: A Personal Account of the Discovery of the Structure of DNA by Nobel Laureate Dr James D Watson. As I was writing, however, I realized that the two books would be quite different in scope and nature. Dr Watson’s book tells of one of the greatest scientific discoveries in the twentieth century. This book is more about an engineering quest than a scientific discovery. By nature, an engineering quest covers a richer fabric of life. The technical ideas behind the quest have first to be discovered and then carried to their logical completion. The discovery may come from luck or inspiration, but the rest of the quest is a matter of perspiration and perseverance. The scope of the book, therefore, is closer to, say, Chuck Yeager’s The Quest for Mach One, his account of breaking the sound barrier. This book depicts the quest to exceed, even if only briefly, the chess-playing ability of the best human chess player on the planet.

相比之下,这本书或许比上面提到的两本书都轻松一些。造成这种情况的一个原因是,创造一台可以击败国际象棋世界冠军的计算机的想法看似轻率。计算机科学家也是人,他们确实喜欢玩乐。另一个原因是我和我的写作深受另一位最喜欢的人的影响我的书《你肯定在开玩笑,费曼先生:一个好奇人物的历险记》,作者是诺贝尔奖获得者理查德·P·费曼博士。从费曼博士的书中,我了解到他特有的幽默感、抑制不住的对生活的热情以及严肃的看待事物的方式。但更重要的是,我了解到能够嘲笑自己的价值。计算机象棋是一个竞争非常激烈的研究领域——科学竞争就像 DNA 研究领域一样激烈和激烈。与国际象棋世界冠军卡斯帕罗夫的决赛也是一场极其严肃的比赛。在一些非常紧张的情况下,嘲笑自己和不把自己太当回事的能力让我保持理智,首先是科学竞争,然后是与加里的两场比赛。如果你能退一步看大局,生活中最严肃的时刻也可能是最有趣的。

Comparatively, this book is perhaps more light-hearted than either of those mentioned above. One reason for this is the seemingly frivolous nature of the very idea of creating a computer that can defeat the World Chess Champion. Computer scientists are humans too, and they do like to have fun. The other reason is that I, and my writing, were greatly influenced by another favorite book of mine, Surely You’re Joking, Mr Feynman: Adventures of a Curious Character by Nobel Laureate Dr Richard P Feynman. From Dr Feynman’s book, I learned of his peculiar sense of humor, irrepressible passion for life, and no-nonsense way of looking at things. But, more importantly, I learned of the value of being able to laugh at myself. Computer chess was a highly competitive research field—scientific rivalry was as strong and heated as in the DNA research area. The final competition with Garry Kasparov, the World Chess Champion, was also an extremely serious affair. The ability to laugh at myself and not to take myself too seriously kept me sane during some very tense situations, arising first from the scientific rivalry and then from the two matches with Garry. If you can take a step back and see the larger picture, the most serious moments in life can also be the funniest.

许多书籍都介绍了国际象棋世界冠军加里·卡斯帕罗夫和深蓝之间的比赛,但其中大多数都集中在“人与机器”方面。这些书也往往是即时国际象棋书籍。“人与机器”的角度显然对于国际象棋书籍来说很畅销,但它并没有抓住比赛的真正本质。这场竞赛实际上是在扮演两种不同角色的男人之间进行的:作为表演者的男人和作为工具制造者的男人。加里和深蓝进行了两场比赛,结果不同。在1996年的比赛中,曼以表演者的身份获胜,在1997年的复赛中,曼以工具制造者的身份获胜。

Many books cover the matches between World Chess Champion Garry Kasparov and Deep Blue, but most of them dwell on the “man versus machine” aspect. These books also tend to be instant chess books. The “man versus machine” angle apparently sells well for chess books, but it does not capture the true essence of the contest. The contest was really between men in two different roles: man as a performer and man as a toolmaker. Two matches were played between Garry and Deep Blue, with different results. In the 1996 match, man won as a performer, and in the 1997 rematch, man won as a toolmaker.

这本书记录了工具制造者和制造机器的人的故事。然而,第二章、第三章的前半部分和附录 A是关于我在项目启动之前的生活的。第二章和第三章前半部分的故事与该项目直接相关。该项目的存在只是因为这些故事中的事件发生了。附录A包含我早年在台湾生活的故事,可以作为背景材料阅读。我放入此材料有两个原因。一是回答什么样的国际象棋低能者会决定与世界上最伟大的国际象棋头脑较量的问题。另一个原因是为您提供有关事情最终为何如此的线索。

This book records the stories of the toolmakers, the people who built the machines. Chapter two, the early part of Chapter three and Appendix A are, however, about my life prior to the launching of the project. The stories in Chapter two and the early part of Chapter three have direct relevance to the project. The project existed only because events in these stories took place. Appendix A contains stories of my early life in Taiwan, and can be read as background material. I put this material in for two reasons. One is to answer the question of what kind of chess imbecile would decide to take on the world’s greatest chess mind. The other reason is to provide you with clues about why things ended up the way they did.

书中的故事跨越两个社会和文化不同的组织:卡内基梅隆大学计算机科学系和 IBM 研究中心。本书的前半部分重点介绍卡内基梅隆大学的项目。在此期间基本技术思路形成,新的计算机国际象棋团队组建一场激烈的科学竞争爆发了。本书的其余部分记录了 IBM 项目的演变:适应新环境、痛苦的人事变动、新国际象棋机器的开发,当然还有加里和深蓝之间的两场激烈且有争议的比赛。这本书的尾声涵盖了赛后事件,包括加里明显避免重赛,尽管在输给深蓝后的两年里,他不断对这样的比赛提出挑战。2000 年 11 月,加里 (Garry) 在世界锦标赛上输给了弗拉基米尔·克拉姆尼克 (Vladimir Kramnik)。这是否有可能与克拉姆尼克进行比赛?目前尚不清楚。

The stories in the book span two socially and culturally different organizations, the Carnegie Mellon Computer Science Department and IBM Research. The first half of the book focuses on the project at Carnegie Mellon. During this period the basic technical ideas took shape, a new computer chess team was formed as a lark, and an intense scientific rivalry erupted. The rest of the book documents the evolution of the project at IBM: adjustment to the new environments, painful personnel changes, development of new chess machines and, of course, two intense and controversial matches between Garry and Deep Blue. The epilogue of the book covers post-match events, including Garry’s apparent avoidance of a replay even though for two years after his defeat by Deep Blue, he had continuously thrown out a challenge for such a match. Garry lost the World Championship to Vladimir Kramnik in November 2000. Will this make a match against Kramnik a possibility? It is not clear at this time of writing.

关于深蓝的一个常见问题是“它智能吗?” 加里在 1997 年比赛期间和之后作弊的指控证实深蓝通过了国际象棋版本的图灵测试(一种盲测,判断你是在与人类还是计算机交互)。但深蓝并不聪明。它只是一个精心制作的工具,在有限的领域中表现出智能行为。加里输掉了比赛,但他才是真正有智慧的选手——深蓝永远不可能提出那些凭空想象的指控。

A frequently-asked question about Deep Blue is, “Is it intelligent?” Garry’s accusations of cheating both during and after the 1997 match confirmed that Deep Blue passed the chess version of the Turing Test (a blind test to tell whether you are interacting with a human or a computer). But Deep Blue is not intelligent. It is only a finely-crafted tool that exhibits intelligent behavior in a limited domain. Garry lost the match, but he was the player with the real intelligence—Deep Blue would never be able to come up with the imaginative accusations.

最后一点:深蓝是团队努力的成果。以第一人称视角描述团队工作的危险之一是,团队其他成员投入的时间和精力可能会被低估。为了减少这个问题,我征求了团队成员对早期草稿的反馈。如果我仍然没有正确地认可某人的贡献,那一切都是我的错。

One final note: Deep Blue was a team effort. One of the dangers of writing a first-person account of a team effort is that the time and the energy put in by other members of the team can be under-represented. To reduce this problem, I have sought out team members’ feedback on early drafts. If I still fail to credit someone’s contribution properly, the fault is all mine.

 

 

致谢

Acknowledgments


贝尔实验室的 Joe Condon 和 Ken Thompson 在国际象棋机器 Belle 上的开创性工作是该项目的直接灵感。肯还提供了深蓝在与加里·卡斯帕罗夫的两场比赛中使用的一些残局数据库。刘易斯·斯蒂勒提供了其他残局数据库。

The pioneering work of Joe Condon and Ken Thompson at Bell Labs, on the chess machine Belle, was the direct inspiration for the project. Ken also provided some of the endgame databases used by Deep Blue in both matches against Garry Kasparov. Lewis Stiller provided the other endgame databases.

卡内基梅隆大学计算机科学系和 IBM 研究中心是使该项目成为可能的两个组织。我要特别感谢这两个组织的许多其他人:

The Computer Science Department at Carnegie Mellon University and IBM Research were the two organizations that made the project possible. There are many other people from both organizations to whom I want to give special thanks:

在卡内基梅隆大学,

At Carnegie Mellon,

• Deep Thought 团队成员Thomas Anantharaman、Mike Browne、Murray Campbell、Peter Jansen 和Andreas Nowatzyk 的贡献以及对我有时不合理行为的容忍。

•    Deep Thought team members Thomas Anantharaman, Mike Browne, Murray Campbell, Peter Jansen, and Andreas Nowatzyk for their contributions and tolerance of my sometimes irrational behavior.

• 我的导师,HT Kung 教授,感谢他的理解和鼓励,并为建立 Deep Thought 提供了资金。

•    My advisor, Professor H T Kung, for his understanding and encouragement and for providing the funding to build Deep Thought.

• Raj Reddy 教授和Randy Bryant 教授在Deep Thought 项目的后期阶段提供了额外的财政援助。

•    Professor Raj Reddy and Professor Randy Bryant, for providing additional financial assistance in the later phases of the Deep Thought project.

• Lawrence Butcher 和John Zsarnay,提供了急需的技术支持。

•    Lawrence Butcher and John Zsarnay, for providing much-needed technical support.

• Hitech 团队的各个成员,特别是Carl Ebeling 和Gordon Goetsch,提供了宝贵的提示和有趣的讨论。汉斯·柏林纳博士(Dr Hans Berliner)更多地是王室反对派,他为我们提供了前进的动力。

•    Various members of the Hitech team, but in particular Carl Ebeling and Gordon Goetsch, for valuable hints and interesting discussions. Dr Hans Berliner, who was more the royal opposition, provided the spur that kept us running forward.

• 其他提供帮助和鼓励的人包括Roberto Bisiani 教授、李开复教授、Tom Mitchell 教授、Danny Sleator 教授、Bob Sproull 教授和Hide Tokuda 教授。

•    Other people who provided help and encouragement include Professor Roberto Bisiani, Kai-Fu Lee, Professor Tom Mitchell, Professor Danny Sleator, Professor Bob Sproull, and Professor Hide Tokuda.

在 IBM 研究中心,

At IBM Research,

• 深蓝团队成员Murray Campbell 和Joe Hoane,感谢他们对该项目的贡献和承诺。顺便,我要感谢默里的妻子吉娜和乔的妻子伊丽莎白,感谢她们对长时间工作和失去的周末的宽容和理解。有几次,默里和乔不得不以他们家人的生命为代价,来帮助我摆脱与深蓝国际象棋筹码相关的一些技术问题。

•    Deep Blue team members Murray Campbell and Joe Hoane, for their contribution and commitment to the project. In passing, I want to thank Murray’s wife Gina and Joe’s wife Elizabeth for their tolerance and understanding of the long work hours and the lost weekends. Several times, Murray and Joe had to bail me out of some technical troubles associated with the bringing up of the Deep Blue chess chips, at the cost of their family lives.

• 整个管理链,尤其是我们的两位经理 Randy Moulic 和 CJ Tan,感谢他们的支持和鼓励。

•    The entire management chain, but especially our two managers, Randy Moulic and C J Tan, for their support and encouragement.

• Jerry Brody,为团队提供技术支持。

•    Jerry Brody, for providing the technical support for the team.

许多外人都提供了帮助。回到 Deep Thought 时代,拉里·考夫曼 (Larry Kaufman)、斯图尔特·克拉克拉夫特 (Stuart Cracraft)、吉姆·吉洛利 (Jim Gillogly)、弗雷德里克·弗里德尔 (Frederic Friedel)、唐·马多克斯 (Don Maddox) 和达普·哈特曼 (Dap Hartmann) 都提供了帮助。IBM 研究中心内外的许多 IBM 员工都参与了这两场深蓝比赛的主办或提供技术支持,ACM 和 TSI 的人员也是如此。我感谢他们所有人。

Many outsiders have helped. Back in Deep Thought days, Larry Kaufman, Stuart Cracraft, Jim Gillogly, Frederic Friedel, Don Maddox and Dap Hartmann all provided help. Numerous IBMers, both in and outside IBM Research, were involved in hosting or providing technical support for the two Deep Blue matches, as did people from ACM and TSI. I thank them all.

如果没有许多大师和其他国际象棋棋手的帮助,深蓝不会取得今天的成就,他们要么与深蓝及其前辈对弈,要么与深蓝及其前辈一起训练。马克西姆·德鲁吉 (Maxim Dlugy)、乔尔·本杰明 (Joel Benjamin)、米格尔·伊列斯卡斯 (Miguel Illescas)、尼克·德弗米安 (Nick Dermian) 和约翰·费多洛维奇 (John Fedorowicz) 等曾一度负有直接责任的特级大师。当然,乔尔也是我们两场深蓝比赛的国际象棋顾问。《纽约时报》国际象棋专栏作家罗伯特·伯恩大师是我的好私人朋友,偶尔会过来玩几盘棋。除了参与该项目的大师之外,罗伯特是提供最多意见的人。最后,如果没有国际象棋世界冠军加里·卡斯帕罗夫的参与,我们永远无法发现深蓝的真正实力。

Deep Blue would not be where it is without the help of many Grandmasters and other chess players who either played games against or trained with Deep Blue and its predecessors. Among the Grandmasters who had direct responsibilities at one time or another were Maxim Dlugy, Joel Benjamin, Miguel Illescas, Nick DeFirmian, and John Fedorowicz. Joel, of course, was also our chess advisor for both Deep Blue matches. Grandmaster Robert Byrne, the New York Times chess columnist, has been a good personal friend who dropped in occasionally for a few games. Besides the Grandmasters who worked on the project, Robert was the one who provided the most input. Finally, without the participation of Garry Kasparov, the World Chess Champion, we would never have found out Deep Blue’s real strength.

许多人审阅了本书的早期草稿并提出了改进建议。丹尼斯·艾利森、托马斯安南塔拉曼, 迈克·布朗, 罗伯特·伯恩, 默里·坎贝尔, Yih-Farn Robin Chen, Joe Hoane, 罗伯特·凯悦, 霍华德·兰德曼, 吉姆·洛伊, 蒂莫西·曼, 安德烈亚斯·诺瓦齐克, 乔治·保罗, 乔纳森·谢弗, 丹尼·斯利特, CJ Tan, 肯·汤普森,和一位匿名审稿人都提供了宝贵的反馈。我的经纪人威廉·克拉克(William Clark)从他独特的角度提出了另一种观点。

Many people have reviewed earlier drafts of this book and made suggestions for improvements. Dennis Allison, Thomas Anantharaman, Mike Browne, Robert Byrne, Murray Campbell, Yih-Farn Robin Chen, Joe Hoane, Robert Hyatt, Howard Landman, Jim Loy, Timothy Mann, Andreas Nowatzyk, George Paul, Jonathan Schaeffer, Danny Sleator, C J Tan, Ken Thompson, and an anonymous reviewer all provided valuable feedback. My agent, William Clark, contributed an alternative view from his unique perspective.

书中的一些照片来自我的私人收藏。许多比赛照片都来自 IBM 收藏。蒙蒂·纽伯恩 (Monty Newborn)、安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 和丹尼·斯利特 (Danny Sleator) 提供了平衡。

Some of the photos in the book are from my private collection. Many of the match photos come from the IBM collection. Monty Newborn, Andreas Nowatzyk, and Danny Sleator provided the balance.

最后但并非最不重要的一点是,我要感谢我的编辑 David Ireland,他完善和润色了我的写作。但最终,任何错误和遗漏均由我承担。

Last but not least, I would like to thank David Ireland, my editor, for refining and polishing my writing. In the end, however, any mistakes and omissions are solely my responsibility.

 

 

国际象棋记谱法

Chess Notation


这不是一本国际象棋书——这里没有真正的国际象棋分析。我尽量避免使用国际象棋符号或国际象棋图表。然而,一些争议,尤其是第二场深蓝比赛中的争议,集中在深蓝的棋局上。理论上,你可以简单地把国际象棋的走法当作标签来追踪争议,但了解这些走法的真正含义也没什么坏处。关于两场深蓝比赛的章节包含国际象棋图表,让您更容易想象比赛情况 - 您可以将它们视为棋盘的照片。背页是本书中使用的代数国际象棋符号的简短描述。

This is not a chess book—there is no real chess analysis here. I have tried to avoid using chess notations or chess diagrams as much as possible. However, some of the controversies, especially those from the second Deep Blue match, centered on the chess moves played by Deep Blue. In theory, you could follow the controversies simply by treating the chess moves as labels, but it does no harm to know what the moves really mean. The chapters on the two Deep Blue matches contain chess diagrams to make it easier for you to visualize the game situations—you can treat them as though they are photos of the chessboard. Overleaf is a short description of the algebraic chess notation used in the book.

该图显示了棋盘的八行(也称为排名)和八列(也称为文件)的名称。这些行从底行到顶部标记为 1 到 8。这些列从最左列到最右列标记为 a 到 h。棋盘上每个方格的名称基于其所在的列和行。因此,棋盘右下角的方格称为h1,左上角的方格称为a8。棋盘本身可以垂直分为两个相等的两半:后侧(a、b、c 和 d 列)和王侧(e、f、g 和 h 列)。两半的名称来自于起始棋盘上国王和王后的位置。

The diagram shows the names for the eight rows (also known as ranks) and the eight columns (also known as files) of the chessboard. The rows are labeled 1 to 8 from the bottom row to the top. The columns are labeled a to h from the leftmost column to the rightmost. The name of each square of the chessboard is based on the column and the row that it is on. Therefore, the lower right corner square of the chessboard is known as h1, and the upper left corner square is known as a8. The chessboard itself can be divided vertically into two equal halves: the queenside (a, b, c, and d file) and the kingside (e, f, g, and h file). The names for the two halves come from the position of kings and queens on the starting chessboard.

不包括兵,棋子有五种:马、象、车、后、王。这些部件可以是白色或黑色。在起始棋盘上,白棋位于第二行,黑棋位于第七行。第一排和第八排的棋子从一排到一排分别是:车、马、主教、王后、国王、主教、马和车。第一级棋子是白色的,第八级棋子是黑色的。持有白棋的玩家称为白棋,持有黑色棋子的玩家称为黑棋。

Not counting the pawn, there are five types of pieces: knight, bishop, rook, queen, and king. The pieces can be colored white or black. On the starting chessboard, white pawns are on the second rank and black pawns are on the seventh rank. The pieces on the first and the eighth rank are, from a file to h file, rook, knight, bishop, queen, king, bishop, knight, and rook. The first rank pieces are white and the eighth rank pieces are black. The player with the white pieces is referred to as White, and the player with the black pieces, Black.

起始棋盘,带有标记的坐标。

The starting chessboard, with labeled coordinates.

棋子移动由棋子移动的目标方格表示。e4 棋是棋子移动到 e4 方格。Pawn 捕获由 pawn 来源的文件的名称表示,后跟目标方格。可选符号“x”可以放置在文件名和方块名称之间。因此,棋子捕获 axb4 使用 a-file 棋子捕获 b4 方格上的棋子。

Pawn moves are represented by the destination square of the pawn move. The move e4 is a pawn move to the e4 square. Pawn captures are represented by the name of the file that the pawn came from, followed by the destination square. An optional symbol ‘x’ can be placed between the file name and square name. Thus, the pawn capture axb4 uses an a-file pawn to capture a piece on the b4 square.

骑士的移动由棋子符号“N”后跟目标方格表示。Nf3 棋步是将马放置在 f3 方格上的棋步。当有不止一名骑士可以移动到同一目标方格时,可以在目标方格名称之前添加源文件的名称或源等级的名称。移动 N2f3 意味着将第二排的马移动到 f3 方格。如果骑士的移动是捕获,则可以在目标方格之前添加可选的捕获符号“x”。因此,Nxg5 棋步是在 g5 方格上吃子的马棋棋步。

Knight moves are represented by the piece symbol ‘N’ followed by the destination square. The move Nf3 is a move that places a knight on the f3 square. When there is more than one knight that can move to the same destination square, the name of the from-file or the name of the from-rank might be added before the name of the destination square. The move N2f3 would mean moving the knight on the second rank to the f3 square. If the knight move is a capture, then the optional capture symbol ‘x’ could be added before the destination square. The move Nxg5 therefore is a knight move that captures a piece on the g5 square.

随后是象走、车走、后走、王走与马的走法相同,但棋子符号分别替换为“B”、“R”、“Q”和“K”。因此,走 Qa4 意味着将皇后移动到 a4 方格。

Bishop moves, rook moves, queen moves, and king moves follow the same format as the knight moves but with the piece symbol replaced by ‘B’, ‘R’, ‘Q’ and ‘K’ respectively. The move Qa4 therefore means moving the queen to the a4 square.

棋步后面的符号“+”表示棋步是过牌;Rc4+ 意味着车到 c4 并过牌。

The symbol ‘+’ following a move means the move is a check; Rc4+ means rook to c4 and giving a check.

棋子升级由棋子的移动和棋子升级到的棋子的符号来表示。因此g8Q意味着g8的典当并晋升为皇后。

Pawn promotions are represented by the pawn’s move followed by the symbol for the piece to which the pawn is promoted. Therefore g8Q means pawn to g8 and promoted to a queen.

易车易位以OO和OOO表示,分别表示王侧易位和后侧易位。

Castling moves are represented by O-O and O-O-O, which mean castling kingside and castling queenside respectively.

移动注释符号'!'、'!!'、'?'、'??' 分别表示“良好的举动”、“出色的举动”、“可疑的举动”和“可疑的举动”。所以h5!意思是“pawn to h5,好棋”。

The move annotation symbols ‘!’, ‘!!’, ‘?’, ‘??’ mean “a good move”, “an excellent move”, “a questionable move” and “a dubious move” respectively. So h5! means “pawn to h5, a good move”.

 

 

序幕

Prologue


 

 

第1章

CHAPTER 1

开演时间!

Show Time!


1997 年 4 月下旬,纽约街头出现了一场不同寻常的国际象棋比赛的海报。画面中,一位 30 岁出头的绅士,面色忧郁、沉思,透过国际象棋注视着观众。他下巴下的小标题写着:“如何让计算机闪烁?” 海报上的绅士是国际象棋世界冠军加里·卡斯帕罗夫(Garry Kasparov),他可能是有史以来最强的国际象棋棋手。

In late April 1997, posters for an unusual chess event were appearing on the streets of New York. They showed a somber and pondering gentleman in his early 30s peering over a chess set at the viewers. The small caption under his chin said, “How do you make a computer blink?” The gentleman on the poster was the World Chess Champion Garry Kasparov, possibly the strongest chess player who has ever lived.

在街边公平大厦的地下室里,我盯着空荡荡的礼堂里的空白屏幕。再过几天,礼堂里就会挤满了人。电视摄像机将被安置在有利的位置,前面的三个巨大的投影屏幕将变得栩栩如生。左侧屏幕将显示大楼 35 层电视演播室的实时图像,该演播室用作游戏室。现场图像通常会显示两名参赛者坐在专门设计的比赛桌上。左边的参赛者是加里·卡斯帕罗夫。另一边的参赛者将是我的两个同事默里·坎贝尔和乔·霍恩中的一个,或者是我。加里真正的对手是我们三个人设计和编程的国际象棋计算机“深蓝”。在游戏过程中,我们只是作为深蓝的延伸,在物理棋盘上为它走棋。在礼堂内,三名国际象棋评论员(有时还有一两名客座评论员)将使用中央屏幕来展示他们对正在进行的比赛的分析。右侧屏幕将显示棋盘的俯视图。这样,观众席上的观众就能清楚地看到当前的比赛局面。

Off the street, in the basement of the Equitable Building, I was staring at the blank screens in an empty auditorium. In a few days, the auditorium would be filled with an overflowing crowd; TV cameras would be entrenched at vantage locations and the three huge projection screens at the front would come to life. The left screen would be showing a live image from a TV studio on the 35th floor of the building, serving as the game room. The live image would usually show the two contestants sitting across a specially designed playing table. The contestant on the left would be Garry Kasparov. The contestant on the other side would be one of my two colleagues, Murray Campbell and Joe Hoane, or me. Garry’s real opponent was the chess computer, Deep Blue, that the three of us had designed and programmed. During the games we acted merely as extensions of Deep Blue and made moves for it on the physical chessboard. In the auditorium itself, three chess commentators, sometimes with a guest commentator or two, would be using the center screen to show their analysis of the ongoing game. The right screen would be displaying the overhead shot of the chessboard. This way, the audience in the auditorium would have a clear view of the present game position.

我花了将近十二年才达到这一点。什么时候我一开始就知道,加里不是世界冠军;他是世界冠军。距离他加冕还有几个月。自 1986 年以来的过去十一年里,我和我的伙伴们一直在不断构建更强大的国际象棋计算机。我们的最终目标是击败世界象棋冠军,无论他或她是谁。

It had taken me almost twelve years to reach this point. When I started, Garry was not the World Champion; it was a few months yet before he was crowned. For the past eleven years, since 1986, my partners and I had been building successively more powerful chess computers. Our eventual goal was to beat the World Chess Champion, whoever he or she was.

在我们之前,许多先驱者,有些是著名的,有些是不太出名的,都为“计算机国际象棋问题”做出了贡献。1949年,克劳德·香农(Claude Shannon)提出了如何对计算机下国际象棋进行编程的建议。从那时起,成千上万的计算机科学家、工程师、业余爱好者、国际象棋棋手,甚至商业组织都致力于解决这个问题。有些人想用国际象棋作为实验工具来了解人类智力是如何运作的。艾伦·纽厄尔、克利夫·肖和赫伯特·西蒙在一篇早期的计算机国际象棋论文中表示:“如果一个人能够设计出一台成功的国际象棋机器,那么它似乎已经渗透到了人类智力活动的核心。” 其他人则将国际象棋视为复杂问题的一个清晰、明确的例子。可以想象,“解决”国际象棋可以提供解决其他复杂问题的新技术。当然,商业实体是为了盈利,而有些人,尤其是业余爱好者,只是为了好玩。

Before us, many pioneers, some famous and some not so famous, had made their contributions to the “Computer Chess Problem”. In 1949, Claude Shannon made his proposal on how to program a computer to play chess. Since then, thousands of computer scientists, engineers, hobbyists, chess players, and even commercial organizations had worked on the problem. Some wanted to use chess as an experimental tool to find out how human intelligence worked. “If one could devise a successful chess machine, one would seem to have penetrated to the core of human intellectual endeavor,” said Allen Newell, Cliff Shaw and Herbert Simon in one of the early computer chess papers. Other people viewed chess as a clear-cut, well-defined example of a complex problem. “Solving” chess could conceivably provide new techniques to solve other complex problems. The commercial entities did it for profit, of course, and some people, especially the hobbyists, did it just for fun.

我们从不同的方向解决这个问题。我们,或者至少是我,将这个问题视为纯粹的工程问题。自 20 世纪 70 年代末以来,人们已经确定,国际象棋计算机随着硬件速度的提高而变得更加强大。到 1985 年,当我开始我的小项目(最终成为深蓝)时,从实验数据推断表明,硬件速度提高一千倍可能足以生产世界冠军级的国际象棋机器。我们的项目始于一个简单的目标,即找出硬件速度的大幅提高是否足以“解决”计算机国际象棋问题。建造这个“所有国际象棋机器之母”本身就是一个有趣的问题。当然,如果我们的机器确实能够击败世界冠军,那将是额外的奖励。

We approached the problem from a different direction. We, or at least I, viewed the problem as a purely engineering one. Since the late 1970s, it had been established that chess computers became stronger as their hardware speed increased. By 1985, when I started my small project that eventually become Deep Blue, the extrapolation from the experimental data indicated that a one thousandfold increase in hardware speed might be sufficient to produce a World Champion-class chess machine. Our project began with a simple goal, namely, to find out whether a massive increase in hardware speed would be sufficient to “solve” the Computer Chess Problem. Building this “Mother of all Chess Machines” was an interesting problem by itself. Of course, it would be an added bonus if our machine could indeed defeat the World Champion.

深蓝的前一个版本于 1996 年在费城输给了加里·卡斯帕罗夫 (Garry Kasparov)。但在那场比赛进行到三分之二时,我们已经与卡斯帕罗夫打成了平手。那个旧版本的深蓝已经比我在 1985 年猜想的机器快了,但这还不够。解决计算机国际象棋问题不仅仅是增加硬件速度。自从那场比赛之后,我们从头开始重建了深蓝,解决了我们遇到的每一场比赛问题,并让大师们广泛参与我们的准备工作。不知何故,所有这些工作都让我们的国际象棋顾问、美国最好的国际象棋棋手之一乔尔·本杰明大师说:“你知道,有时深蓝也会下棋。” 乔尔已经无法准确地区分深蓝的招式和顶级宗师的招式了。

The previous version of Deep Blue, lost a match to Garry Kasparov in Philadelphia in 1996. But two-thirds of the way into that match, we had played to a tie with Kasparov. That old version of Deep Blue was already faster than the machine that I conjectured in 1985, and yet it was not enough. There was more to solving the Computer Chess Problem than just increasing the hardware speed. Since that match, we rebuilt Deep Blue from scratch, going through every match problem we had and engaging Grandmasters extensively in our preparations. Somehow, all the work caused Grandmaster Joel Benjamin, our chess advisor, and one of the best chess players in the US, to say, “You know, sometimes Deep Blue plays chess.” Joel could no longer distinguish with certainty Deep Blue’s moves from the moves played by the top Grandmasters.

媒体对这场新比赛充满期待。如果新款深蓝赢得了这场比赛,那么这将是人类作为工具制造者的悠久历史上的一个重要时刻。这也将是计算机科学家和人工智能研究人员期待已久的里程碑的完成。几乎可以肯定,这场比赛将比任何世界象棋锦标赛的比赛规模更大,唯一的例外可能是 1972 年费舍尔对斯帕斯基的比赛。如果我们真的赢了,也许连费舍尔对斯帕斯基的比赛也无法与之相比。

The press covered this new match with much anticipation. If the new Deep Blue won the match, then it would be a momentous occasion in the long history of men as toolmakers. It would also be the completion of a long-sought-after milestone for computer scientists and artificial intelligence researchers. It was almost certain that this match would be bigger than any World Chess Championship match, with possibly the sole exception of the Fischer vs. Spassky match in 1972. If we did win, perhaps not even that Fischer vs. Spassky match would compare.

新的深蓝有了很大的改进,但这就足够了吗?我和伙伴们多年前开始的旅程终于要结束了吗?

The new Deep Blue was much improved, but would it be enough? Would the journey begun by my partners and me so many years ago finally be over?

 

 

卡内基梅隆大学

Carnegie Mellon


 

 

第2章

CHAPTER 2

麻烦制造者办公室

An Office of Troublemakers


“你的办公室是麻烦制造者的办公室!” 孔一脸无奈的说道。Kung 就是 HT Kung 教授,他是我在匹兹堡卡内基梅隆大学的导师。他是一个四十多岁的高个子,体格健壮。我是计算机科学(CS)系的研究生。我们刚刚在他的办公室进行了一次长时间的私人谈话。

“Your office is an office of troublemakers!” Kung exclaimed, with a resigned look. Kung was Professor H T Kung, my faculty advisor at Carnegie Mellon University in Pittsburgh. He was a tall, well-built man in his 40s. I was a graduate student in the Computer Science (CS) Department. We had just had a long private conversation in his office.

那是1987年的秋天。我在卡内基梅隆大学期间,孔教授一直是我的导师。我非常尊重他,不仅因为他是我的好顾问,还因为他私下所做的事。孔教授是孔子的嫡系后裔,他为尽自己所能帮助世界各地华人社会的技术和经济发展而感到自豪。我和孔教授没有任何矛盾。我的麻烦出在系里的另一位教员身上。

It was the autumn of 1987. Professor Kung had been my advisor during my entire stay at Carnegie Mellon. I had a great deal of respect for him, not just because he was a good advisor to me, but also because of the deeds that he did in private. Professor Kung was a direct descendent of Confucius and took great pride in doing whatever he could to assist the technical and economic developments of Chinese societies around the world. I was not having trouble with Professor Kung. My troubles were with another faculty member in the department.

谈话时,卡内基梅隆大学有两个计算机国际象棋队。Hans Berliner 博士领导着 Hitech 团队。ChipTest 的竞争团队是一群自由意志、大多数无人监督的研究生,我就是其中之一。我得到了 Kung 教授的一些指导,而我的团队成员在 ChipTest 上工作几乎纯粹是为了好玩。两队的比赛原本是友好的,但即将演变成一场激烈的对抗。

At the time of the conversation, there were two computer chess teams at Carnegie Mellon. Dr Hans Berliner headed the Hitech team. The competing ChipTest team was a group of free-willed, mostly unsupervised graduate students, of which I was one. I had some light faculty supervision from Professor Kung, while my team members worked on ChipTest almost purely for fun. The competition between the two teams had been friendly, but it was about to turn into an intense rivalry.

事后看来,竞争是不可避免的。两队之间曾发生过一些摩擦。ChipTest 团队对经常出现的不正确的公开描述感到不满,即柏林博士正在监督我们或我们是他的学生。在很大程度上,这种描述是柏林纳博士在计算机领域现有声誉的结果国际象棋,这让记者自动假设1他领导了卡内基梅隆大学所有的计算机国际象棋工作。这种描述对 ChipTest 团队来说有些侮辱,但仅凭这一点并不会导致争论立即升级。

In hindsight, the rivalry was inevitable. There had been some friction between the two teams. The ChipTest team was not happy with the frequent and incorrect public portrayal that Dr Berliner was supervising us or that we were his students. To a great extent this portrayal was the result of Dr Berliner’s existing fame in computer chess, which led reporters automatically to assume1 that he led all computer chess efforts at Carnegie Mellon. This characterization was somewhat insulting to the ChipTest team, but it alone would not have caused the immediate escalation of the contention.

此次升级的起因是 ChipTest 团队收到参加国际象棋锦标赛的邀请。Hitech 团队碰巧要参加同一场比赛,Berliner 向 Kung 解释了为什么 ChipTest 应该跳过比赛2

The escalation was triggered by, of all things, an invitation for the ChipTest team to play in a chess tournament. The Hitech team happened to be going to the same tournament and Berliner had made his case to Kung on why ChipTest should skip the event2.

我和孔教授的谈话正是关于如何处理国际象棋邀请。Kung 支持 ChipTest 团队,认为我们没有理由不参加比赛。然而,为了避免出现进一步的问题,ChipTest 团队在谈话之前就已经决定尊重 Berliner 博士的建议。

My conversation with Professor Kung was precisely on what to do with the chess invitation. Kung was supportive of the ChipTest team and saw no reason why we should not play. However, to avoid further problems, the ChipTest team had already decided to honor Dr Berliner’s suggestion prior to the conversation.

孔教授见问题似乎解决了,松了口气,但当他说出“麻烦制造者”的言论时,他一定有一种似曾相识的感觉。两年前,即 1985 年,我们进行了另一次谈话,也是关于柏林博士的问题。那次谈话是导致 ChipTest 和最终 Deep Blue 创建的一系列事件的关键部分。

Professor Kung was relieved that the issue seemed to be resolved, but when he made his remark about the “troublemakers”, he must have had a sense of déjà vu. Two years before, in 1985, we had had another conversation, also on an issue related to Dr Berliner. That other conversation was a critical part of the chain of events that led to the creation of ChipTest and, eventually, Deep Blue.

深蓝项目的真正种子可以追溯到更远的地方。这一切都是从孔教授的“麻烦制造者”开始的……

The real seed of the Deep Blue project can be traced back even further. It all began with Professor Kung’s “troublemakers” …

一年级研究生

First Year Graduate Students

卡内基梅隆大学的计算机科学(CS)系是该大学的。其他几个学科都进入了前十名,但计算机科学名列前三,另外两个是麻省理工学院和斯坦福大学的。根据个人喜好,有人甚至会说卡内基梅隆大学的CS系是最好的。该系有自己独特的氛围,研究生在自己的研究方向上比其他学院有更多的发言权。事实上,该系只提供博士学位课程(没有学士或硕士学位课程)可能给了教师们进行实验并提出独特系统的机会。

The Computer Science (CS) Department at Carnegie Mellon was the Department of the university. Several others ranked among the top ten in their disciplines, but CS was ranked among the top three, the other two being the ones at MIT and Stanford. Depending on personal preferences, some would even say that the CS Department at Carnegie Mellon was the best. The Department had its own unique atmosphere, and the graduate students had a great deal more say in their own research directions than at other schools. The fact that the Department offered only a PhD program (no BS or MS program) probably gave the faculty the chance to experiment and come up with the unique system.

部分由于不寻常的制度,部分由于新生的背景不同,每个秋季学期的整个第一个月都被指定为“移民课程”。对于即将入学的学生来说,这是一个了解环境、工具、项目和人员的时间,并且总体上是一个愉快的时光。年长的学生会花时间展示他们的项目,帮助新学生适应,当然,他们自己通常也玩得很开心。移民课程临近结束时,一年一度的部门“招待会”将举行,通常在匹兹堡的某个高档地点。我对作为一名饥饿的研究生在接待处大吃大喝的经历有着美好的回忆。移民课程也是“结婚”过程发生的时间。在移民课程期间,新生们寻找对他们的研究感兴趣的教师,并决定他们想要谁作为他们的教师顾问。当然,教师们也会观察新生,决定是否想要某个特定的学生。与现实生活中一样,婚姻并不是最终的,教师和学生之间的“离婚”有时确实会发生,部分原因是学生的资金来源不一定与特定项目挂钩。教师主要通过受指导者对他们的尊重来控制他们的受指导者。除了课程作业外,教师对非顾问学生的控制力微乎其微。

In part because of the unusual system and in part because of the diverse background of the incoming students, the entire first month of each fall semester was designated the “Immigration Course”. For the incoming students, it was a time to learn about the environments, tools, projects, and people, and generally to have a good time. The older students would spend time demonstrating their projects, helping the new students to adapt, and of course, generally having a good time themselves. Near the end of the Immigration Course, the annual Department “Reception” would be held, usually at some fancy location in Pittsburgh. I have a fond memory of pigging out at the Reception as a starving graduate student. The Immigration Course was also the time the “Marriage” process took place. During the Immigration Course, the new students sought out the faculty whose research interested them, and decided whom they wanted as their faculty advisors. The faculty, of course, also observed the new students, deciding whether they wanted a particular student. As in real life, the Marriages were not final, and “divorces” between faculty members and students did occur sometimes, made possible partly by the fact that the funding sources for the students were not necessarily tied to specific projects. Faculty members controlled their advisees mainly through the respect that their advisees had for them. Other than course work, faculty members had very little hold over non-advisee students.

1982年,也就是我“移民”入系的那一年,录取学生的录取率高于预期,而且出现了名额紧张的情况。通常情况下,在分配办公室时,最好将新生和老学生混合在一起。由于空间紧张,我和另外三名即将入学的学生迈克·布朗、安德烈亚斯·诺瓦齐克和托尼·斯坦茨一起来到了一间空荡荡的办公室。最后,我们四个人都平安无事,但一路上老师们可能希望他们没有这么做。第二年,托尼搬到了另一间办公室,接下来的几年里,只有迈克、安德烈亚斯和我占据了这个办公室,成为了亲密的朋友。孔教授所说的麻烦制造者,很可能不包括托尼。

In 1982, the year I “immigrated” into the Department, the acceptance rate of the admitted students was higher than expected, and there was a space crunch. Normally when offices were assigned, it was considered preferable to mingle the incoming students with older students. As a result of the space crunch I ended up in an empty office together with three other incoming students, Mike Browne, Andreas Nowatzyk and Tony Stentz. In the end, the four of us came out all right, but along the way the faculty probably wished that they had done otherwise. The next year, Tony moved to another office, and for several years only Mike, Andreas and I occupied the office, becoming close friends. Professor Kung’s remark on troublemakers probably did not include Tony.

迈克可能是第一个与我进行真正对话的美国人,这是我在卡内基梅隆大学校园的第一天进行的。这也是我在美国的第三天。当我打开新分配的办公室的门时,我发现里面有一个留着胡子的男人。这是迈克。由于他的胡子,我看不出他有多大,但他是新班里年龄较小的学生之一。迈克出生于新泽西州,但在宾夕法尼亚州阿伦敦长大,仅用三年时间就在里海大学完成了本科学位。在我们的谈话中,迈克将自己描述为“黑客”,他将其定义为可以编写计算机程序的人。也就是说,尽管是在一个专门的领域,但还是有创造的人。这个定义至今仍在使用,尽管媒体使用的更常见的定义是通过虚拟手段损坏或利用计算机系统的人。新的定义描述了具有破坏性的人。迈克显然非常聪明,但他也令人惊讶地谦虚。他将自己的加速大学教育完全归因于他在多项选择测试中取得好成绩的能力。“我只是碰巧擅长多项选择题测试,”他说,这表明他没有其他有用的才能。

Mike was probably the first American with whom I had a real conversation, which took place on my first day on the Carnegie Mellon campus. It was also only my third day in the United States. When I opened the door to my newly-assigned office, I found a bearded man inside. This was Mike. Because of his beard, I could not tell quite how old he was, but he turned out to be one of the younger students in the incoming class. Born in New Jersey, but growing up in Allentown, Pennsylvania, Mike finished his undergraduate degree at Lehigh University in just three years. In our conversation Mike described himself as a “hacker”, which he defined as someone who could churn out computer programs. That is, someone who creates, although in a specialized domain. This definition is still in use today, although the more common definition used by the media is someone who damages or exploits computer systems through virtual means. The newer definition describes a destructive person. Mike was obviously very smart, but he was also surprisingly self-effacing. He attributed his accelerated college schooling entirely to his ability to do well in multiple-choice tests. “I just happened to be good at taking multiple-choice tests,” he said, suggesting that he had no other useful talent.

迈克是年轻的新生之一,而安德烈亚斯和我则是年纪较大的学生。就我而言,这是因为在我的家乡台湾,男孩在大学毕业后必须服兵役两年3。安德烈亚斯年纪更大,因为他多花了一年时间才获得两个艰难的本科学位:物理学和计算机科学。他来自德国,毕业于汉堡大学,在德国攻读博士学位一年后,他决定搬到卡内基梅隆大学。迈克是一位非常优秀的程序员,安德烈亚斯也不是太差劲。安德烈亚斯不仅是一位顶尖的程序员,而且还是一位拥有多种技术技能的人。当我们在卡内基梅隆大学时,他一直给我带来惊喜。我坚信,只要我下定决心,我就能擅长任何技术性的事情。安德烈亚斯似乎已经擅长一切技术。

While Mike was one of the younger incoming students, Andreas and I were among the older ones. In my case, this was because in Taiwan, where I came from, boys were required to serve two years in the military after college 3. Andreas was older because he had spent one extra year to get two tough undergraduate degrees: physics and computer science. He hailed from Germany, was a graduate of the University of Hamburg, and had been one year into his PhD program in Germany before he decided to make the move to Carnegie Mellon. Mike was an extremely good programmer, and Andreas was not too shabby either. Besides being a top-notch programmer, Andreas turned out to be a man with many technical skills. He never ceased to amaze me while we were at Carnegie Mellon. I had a dogged belief that I could be good at anything technical if I put my mind to it. Andreas seemed to be already good at everything technical.

迈克和我在卡内基梅隆大学的早年都有点贪玩,而安德烈亚斯从第一天起就很忙。直到很久以后,当我们发现我们有许多共同的技术兴趣时,安德烈亚斯和我才真正成为亲密的朋友。然而,这种友谊并不是真正对称的。我记得安德烈亚斯在我所在领域帮助过我的很多例子不足。但他是如此自给自足,以至于我不记得曾经做过任何事情来帮助他。

Both Mike and I were somewhat playful in our early years at Carnegie Mellon, while Andreas was all business from day one. Andreas and I did not really become close friends until much later when we found that we shared many common technical interests. The friendship, however, was not really symmetrical. I can remember many instances of Andreas helping me in areas where I was lacking. But he was so self-sufficient that I don’t recall ever doing anything to in any way help him.

在到达卡内基梅隆大学之前,安德烈亚斯、迈克和我都曾接触过国际象棋或电脑象棋。迈克在高中时曾是一位认真的国际象棋锦标赛选手。安德烈亚斯并不认真下棋,但他一直对电脑象棋有着浓厚的兴趣。本科生期间,他曾在德国担任世界微机国际象棋锦标赛的裁判。我早期接触电脑象棋主要是通过阅读技术期刊和书籍。在我们移民期间,安德烈亚斯和我检查了该系现有的计算机国际象棋研究。我们都对超大规模集成 (VLSI) 芯片设计感兴趣,而 VLSI 小组的研究生 Carl Ebeling 正在设计一种基于 VLSI 的国际象棋机。安德烈亚斯和我都没有足够的兴趣对电脑国际象棋做任何事情。当时我认为电脑象棋还有很长的路要走,我不知道它会如何结束。此外,我真的很感兴趣做一些实际的事情。

Andreas, Mike, and I all had some prior contact with chess, or computer chess, before arriving in Carnegie Mellon. Mike used to be a serious tournament chess player when he was in high school. Andreas did not play chess seriously, but he had been keenly interested in computer chess. He had once helped to referee a World Microcomputer Chess Championship tournament in Germany while an undergraduate student. My earlier contacts with computer chess were mainly from reading technical journals and books. During our immigration both Andreas and I checked out the existing computer chess research in the Department. We were both interested in Very Large Scale Integration (VLSI) chip design and Carl Ebeling, a graduate student in the VLSI group, was designing a VLSI-based chess machine. Neither Andreas nor I were sufficiently interested to do anything with computer chess. At the time I thought computer chess still had a long way to go and I did not see how it would end. Moreover, I was really interested in doing something practical.

Kung 教授是 VLSI 小组的主要教员,对 Andreas 和我来说都是显而易见的选择。迈克决定与一位人工智能 (AI) 教授合作。

Professor Kung was the leading faculty member in the VLSI group and the obvious choice for both Andreas and me. Mike decided to work with a professor in Artificial Intelligence (AI).

总体而言,安德烈亚斯对人工智能研究的评价很低。迈克一开始并没有持相同的观点,但在他与第一位顾问分手后,他和安德烈亚斯都将人工智能称为废话。我没有走同样的极端,但我看到了一些所谓的人工智能研究,它们确实配得上狗屎的标签。迄今为止,该系最大的群体是人工智能组,包括几乎一半的学生。我们办公室对人工智能的大多数观点显然是部门内的少数观点。

Andreas had a very low opinion of AI research in general. Mike did not share the same opinion at first but, after he split with his first advisor, both he and Andreas referred to AI as bullshit. I did not go to the same extreme, but I had seen some so-called research in AI that really deserved the bullshit label. The largest group by far in the Department was the AI group, including almost half of the students. The majority view toward AI in our office was clearly the minority view in the Department.

尽管我们对人工智能的评价都不高,但多年来,迈克、安德烈亚斯和我就计算机国际象棋进行了多次对话。其中一次谈话是关于 Belle,20 世纪 80 年代初期的顶级国际象棋机器。当安德烈亚斯还在汉堡大学时,他参加了百丽设计师之一肯·汤普森的演讲。在演讲中,肯描述了一些实验,这些实验表明,当给予更多时间时,贝尔的弹奏能力会更强。我提出了一个问题:如果像贝尔一样的机器速度快一千倍,会发生什么。我们没有人相信这样的机器足以击败世界冠军,但它可能很接近。无论如何,当我们谈话时,没有人我们意识到我会把安德烈亚斯和迈克拉进一个计算机国际象棋项目,试图回答这个问题。

Despite our mutual low opinion of AI, over the years Mike, Andreas and I had several conversations about computer chess. One of the conversations was about Belle, the top chess machine from the early 1980s. While Andreas was still at the University of Hamburg he had attended a talk by Ken Thompson, one of the Belle designers. In the talk, Ken described experiments that showed Belle playing more strongly when given more time. I posed the question of what would happen if a Belle-like machine were, say, one thousand times faster. None of us believed that such a machine would be sufficient to beat the World Champion, but it might be close. Anyway, when we had the conversation, none of us realized that I would drag both Andreas and Mike into a computer chess project trying to answer that very question.

孔教授所说的麻烦制造者或许也包括了安德烈亚斯,因为几年后安德烈亚斯和他闹翻了。我不知道迈克的分裂有多严重,但这可能已经足够糟糕了,以至于孔教授在每学期末举行的“黑色星期五”教职会议上注意到了这一点。安德烈亚斯很可能在孔教授的“麻烦制造者”名单上名列前茅,但我可能位居榜首,因为我是许多其他麻烦的根源。我造成的一些麻烦是由于不成熟的恶作剧。其他麻烦并不完全是我造成的,正如你将在接下来的几章中看到的那样。

Professor Kung’s remark about troublemakers perhaps included Andreas as, after a couple of years, Andreas and he had a falling-out. I don’t know how bad Mike’s split was, but it might have been bad enough for Professor Kung to have noticed at the “Black Friday” faculty meeting held at the end of each semester. Andreas could well have been high on Professor Kung’s list of “troublemakers”, but I probably took the top slot since I was the source of a lot of other troubles. Some of the troubles I caused resulted from immature pranks. Other troubles were not entirely my doing, as you will see in the next few chapters.

我早年的恶作剧产生了两个深远的影响。首先,我在学生中获得了可疑的名声。具有讽刺意味的是,这种声誉有助于招募其他学生参与国际象棋项目。(当然,招募其他人只是第一步。最终,项目本身必须有趣,我必须正确对待我的朋友。)

The pranks that I pulled in the early years had two profound effects. The first was that I gained a reputation of a questionable nature among the students. This reputation ironically was helpful in recruiting other students to work on the chess project. (Of course, recruiting others was just the first step. In the end, the project had to be interesting in its own right and I had to treat my friends properly.)

第二个影响是将我列入该部门的濒危物种名单。我的第一个恶作剧(在下一节中提到)并没有在黑色星期五教员会议上为我赢得任何朋友。我的第二次恶作剧几乎让我被学校开除,如果几年后发生的话,我可能会入狱。

The second effect was to place me on the endangered species list at the Department. My first prank, mentioned in the next section, did not win me any friends at the Black Friday faculty meeting. My second prank almost got me kicked out of the school and could have landed me in jail if it had taken place a few years later.

恶作剧的许可证

A License to Make Mischief

台湾学生应该严格遵守校规,当我还在台湾时,我总是对美国大学校园里学生们巧妙的愚人节恶作剧的故事着迷。

Taiwanese students were supposed to pay strict attention to school rules and, while I was still in Taiwan, I was always fascinated by stories of students pulling ingenious April Fool’s pranks on US college campuses.

在移民课程期间,每个人都参加的课程之一是部门介绍,我们都挤进一个大教室,向部门的其他成员介绍自己。在向即将入学的学生自我介绍后,时任系行政助理4 的莎朗·伯克斯 (Sharon Burks)开玩笑地建议,无论是谁伪造了一份文件,4 月 1 日她发来的电子帖子请留意。听到高年级学生如此不服从命令,我感到很惊讶,我对自己说:“这很好。所以我们有恶作剧的许可。”

During the Immigration Course, one of the sessions attended by everyone was the Department Introduction where we all squeezed into a large classroom and introduced ourselves to the rest of the Department. After introducing herself to the incoming students, Sharon Burks, who was the Department Administration Assistant at the time4, jokingly advised whoever had forged an electronic post from her on April 1 to watch out. Surprised to hear of such insubordination from the older students, I said to myself, “This is neat. So we have a license to make mischief.”

我在卡内基梅隆大学的第一年很忙碌。我想尽快通过所有核心资格赛。在台湾服兵役的两年,让我生疏了,也有很多需要追赶的地方。第一个学期结束后,我在学校的表现还不错,我在卡内基梅隆大学的第一个愚人节即将到来。

My first year at Carnegie Mellon was a busy one. I wanted to pass all the core qualifiers as fast as I could. The two years spent in military service in Taiwan had made me rusty so I also had a lot of ground to catch up. After the first semester, I was in decent shape school-wise, and my first April Fool’s day at Carnegie Mellon was coming up.

早在 20 世纪 80 年代初,即使是在卡内基梅隆大学这样的顶级计算机研究生院,研究生也只能保证能够使用办公室中的纯文本终端。我们中的一些人拥有可以显示黑白图形的办公室工作站,航站楼还有一些其他工作站供公众使用。但我们的大部分工作仍然是在连接到当时最先进的小型计算机的终端上完成的。小型计算机的功能不如我们今天能以不到 300 美元购买的个人微型计算机(顺便说一句,这甚至不足以购买我们使用的终端之一)。大多数小型机运行 UNIX 操作系统5。如今,UNIX 广泛应用于无法容忍操作系统导致的计算机崩溃的领域。

Back in the early 1980s, even in a top CS graduate school like Carnegie Mellon, the graduate students were only guaranteed access to text-only terminals in their offices. Some of us had office workstations that could display black-and-white graphics, and there were a few other workstations for public use in the Terminal Room. But most of our work was still done on terminals connected to then state-of-the-art minicomputers. The minicomputers were less powerful than the personal microcomputers that we can buy today for less than $300 (which, by the way, was not enough to buy even one of the terminals that we used). Most of the minicomputers ran the UNIX operating system 5. UNIX is widely used today in areas where computer crashes caused by the operating system are not tolerable.

UNIX 允许您与同一台计算机上的另一个用户“交谈”,前提是其他用户授予了该权限。大多数用户默认授予此权限。当您与其他用户“交谈”时,UNIX 会将您键入的字符直接写入其他用户的终端上。通常情况下,这不会造成任何伤害,除非你变得令人讨厌,否则可能会惹恼其他用户。然而,当其他用户使用的计算机终端有些“智能”时,有一种方法会造成严重破坏。更智能的计算机终端通常会接受计算机可用来移动光标(屏幕上突出显示新字符将出现的位置的指示器)并读回光标的 (x, y) 位置的命令字符串。通过发送适当的命令字符串,您可以将光标移动到特定的屏幕位置,然后使终端将光标的位置发送回受害者的计算机。由于计算机是很可能期待用户命令,通常看起来好像受害者已将光标位置作为命令输入。转眼间,你已经让你的受害者执行了你自己选择的命令。这就是我 1983 年愚人节恶作剧背后的基本想法。

UNIX allows you to “talk” to another user on the same computer, assuming that the other user grants the privilege. Most users grant this privilege by default. When you “talk” to the other user, UNIX writes your typed characters directly onto the other user’s terminal. Normally this would be harmless, other than possibly annoying the hell out of the other user if you get obnoxious. There is, however, a way to wreak havoc when the computer terminal used by the other user is somewhat “intelligent”. The smarter computer terminals usually will accept command strings that the computer can use to move the cursor (the indicator on the screen which highlights where the new character is to appear) and to read back the (x, y) location of the cursor. By sending appropriate command strings, you can move the cursor to a particular screen location and then cause the terminal to send out the location of the cursor back to your victim’s computer. Since the computer is very likely expecting a user command, it usually appears as if the victim has entered the cursor location as a command. Presto, you have caused your victim to execute a command of your own choosing. This was the basic idea behind my April Fool’s prank in 1983.

3月31日,我和一位同事进行了交谈,然后他就介入了。我们的计划如下:我要编写两个程序。第一个程序将随机选择受害者并向他们的终端发送适当的字符串。然后,这些字符串将导致受害者的终端发出执行我的第二个程序的命令。第二个程序会从受害者的帐户中在部门电子公告板上随机发布一条可能不适当的消息。我的同事负责为帖子收集随机消息,他做得很好。整个事情中最困难的技术部分是找到一种方法,让受害者即使在终端前也不知道发生了什么。光标必须移动到适当的位置以建立计算机命令来执行第二个程序,然后必须返回到其原始位置。整个过程结束后,屏幕看起来应该与攻击前一模一样。这很棘手,但可行。即使坐在终端前,受害者也只能看到类似电源闪烁的东西,光标闪烁大约一秒钟。直到 4 月 1 日晚上,这些节目才准备就绪,我们在当天剩下的几个小时里让它们自由发挥。

On March 31, I had a conversation with a colleague who then got involved. Our plan was the following: I was to write two programs. The first program would randomly select victims and send out appropriate strings to their terminals. The strings would then cause the victim’s terminal to send out a command that executed my second program. The second program would post a random, possibly inappropriate, message on the Department electronic bulletin board from the victim’s account. My colleague was to come up with the collection of random messages for the posts, and he did a great job. The hardest technical part of the whole thing was to figure out a way that the victims would not know what had happened, even when they were in front of their terminals. The cursor had to be moved to the appropriate locations to build up the computer command to execute the second program, and it then had to get back to its original location. After the whole process, the screen should look exactly the same as before the attack. It was tricky, but doable. Even sitting in front of the terminal, the victim would only see something similar to a power flicker with the cursor flashing around for about a second. The programs were not ready until the evening of April 1, and we let them loose for the remaining few hours of the day.

第二天早上,我们看到电子公告板上出现了几条不寻常的信息,其中大多数受害者是教职员工6。当时,一名教职员工收到一条有趣的信息,询问他在女厕所里发现的一条连裤袜的主人是谁。另一位教员“发帖”称,自己成功登录电脑,但不知道如何注销,需要帮助。他并没有被逗乐,也没有意识到今天是什么日子,他真的发布了一条消息,要求发布注销消息的人招供。我恢复了我的程序,但这次只有一个允许的受害者,即我,并且消息文件被修改为仅包含一条消息。该消息首先指出,教授们有时会忘记,前一天是 4 月 1 日。此外,有明显迹象表明这些消息不是真实的。所有消息标题都用引号括起来,从技术上讲,该消息确实是从教授自己的帐户发布的,就像我的消息是通过相同的机制从我的帐户发布的一样。

The next morning we saw several unusual messages showing up on the electronic bulletin board, with most of the victims being faculty members 6. An interesting message had a faculty member at the time, asking who was the owner of a pair of panty hose that he found in the ladies’ room. Another faculty member “posted” a message saying that he managed to log on to the computer, but did not know how to log off and needed help. He was not amused, and not realizing what day it was, he posted, for real, a message demanding that whoever posted the logging-off message confess. I resurrected my programs, but this time with only one allowed victim, namely me, and the message file was modified to contain only one message. The message first pointed out that professors were sometimes forgetful and that the previous day was April 1. Furthermore, there were clear indications that the messages were not real. All of the message titles were enclosed in quotation marks and, technically, the message was indeed posted from the professor’s own account, just as my message was posted from my account through the same mechanism.

我没有因这起事件而遭受直接后果。但这,以及我的下一个更严重的恶作剧,可能影响了我在一些教职人员眼中的地位。

I suffered no direct consequence from this incident. But this, and my next far more serious prank, probably affected my standing in the eyes of some faculty members.

在此事件之后,我的同事永久禁用了他终端的写入权限。没有人会对他下手!另一位朋友对玩光标控制很感兴趣,并最终创建了一个“迷幻的命令提示符”。每次他输入命令,终端屏幕上就会出现一个牛仔般的身影,而屏幕左侧就会射出一支箭矢,射中牛仔,然后牛仔就化为乌有。

After this episode, my colleague disabled the writing privilege to his terminal permanently. Nobody was ever going to do the trick to him! Another friend got interested in playing with the cursor control and eventually created a “psychedelic command prompt”. Every time he typed in a command a cowboy-like figure would be displayed on his terminal screen, and an arrow would shoot out from the left side of the screen and hit the cowboy who then disintegrated into nothing.

罗杰斯先生的邻居

Mister Rogers’ Neighborhood

UNIX是一个相对安全的操作系统。它的创建者之一肯·汤普森在他的图灵奖演讲中提到了几种妥协它的方法。但一般来说,维护良好的 UNIX 系统几乎不会受到业余攻击。问题是大多数 UNIX 系统不一定维护得很好。

UNIX is a relatively secure operating system. Ken Thompson, one of its creators, mentioned several ways to compromise it in his Turing Award Lecture. But generally, a well-maintained UNIX system is almost immune to amateurish attacks. The problem is that most UNIX systems are not necessarily well maintained.

在 20 世纪 80 年代初期,大学中几乎所有较新的计算机都运行 UNIX。其中许多系统都连接到 DARPA(国防高级研究计划局)网络,即当今互联网的前身。DARPA 网络经常被缩写为 Arpanet。由 DARPA 赞助的计算机科学系的许多研究项目都在阿帕网上。

In the early 1980s, almost all of the newer computers in universities ran UNIX. Many of these systems were connected to the DARPA (Defense Advanced Research Project Agency) Net, the precursor to today’s Internet. The DARPA net was frequently shortened to the Arpanet. The CS Department, with a lot of research projects sponsored by DARPA, was on the Arpanet.

该部门有专门且有能力的人员维护计算机系统,但我始终无法找到破坏部门主要机器的方法。然而,阿帕网上有许多不太安全的机器。

The Department had a dedicated and capable staff maintaining the computer systems, and I never could figure out a way to break the main departmental machines. Out there on the Arpanet, however, there were lots of less secure machines.

在该部门,所有可从阿帕网访问的计算机帐户都有密码,必须在登录前输入密码。对于网络上的许多计算机来说,情况并非如此。我偶然发现7在加州大学伯克利分校的一台机器上有一个访客帐户没有任何密码。有人把门半开着。现在,仅凭这一行为并不一定意味着存在问题。UNIX操作系统允许你锁定自己的私有财产,使其他用户无法访问你的网上情书。(尽管我不明白为什么你要把情书留在网上。) 根据定义,访客帐户的权利有限,因此访客不应该能够解锁锁。所以一切都好,对吧?

In the Department, all computer accounts accessible from the Arpanet had passwords that had to be entered before logging on. This was not true for many computers on the Net. I stumbled 7 onto a machine at UC Berkeley that had a guest account without any password. Someone left the door ajar. Now this act alone does not necessarily mean there was a problem. The UNIX operating system allows you to lock up your own private property so that other users cannot access your on-line love letters. (Although why you would leave your love letters on line is beyond me.) The guest account, by definition, had limited rights, therefore a guest should not be able to unlock the locks. So everything’s OK, right?

在现实生活中,大楼管理员通常拥有可以打开大楼内任何门的万能钥匙。在UNIX系统上,系统管理员可以成为所谓的“超级用户”,可以对系统执行任何操作。某些维护系统所需的可信程序通常也被授予超级用户权限,以便程序能够正常运行。有时,对这些超级用户程序的信任是错误的。程序本身可能没问题,但它们运行所依赖的文件可能不安全。任何称职的系统管理员都会确保超级用户程序本身不会受到损害,但超级用户程序引用的文件可能会被经验不足的系统管理员忽略。在伯克利的机器上,无论出于何种原因,一些文件没有得到适当的保护,并且有一个后门,客人可以进入并自行获取主密钥。

In real life, the building managers usually have master keys that they can use to open any door in the building. On a UNIX system, the system manager can become a so-called “superuser” who can do anything to the system. Certain trusted programs that are needed to maintain the system are usually also given the superuser privilege so that the programs can function properly. Occasionally, the trust placed in these superuser programs is misplaced. The programs themselves may be OK, but the files that they rely on to function might be insecure. Any competent system manager would make sure that the superuser programs themselves cannot be compromised, but the files referenced by the superuser programs can, and do, get neglected by less experienced system managers. On the Berkeley machine, for whatever reason, some of the files were not properly secured, and there was a backdoor where a guest could get in and help himself to a master key.

于是我成为了伯克利机器上的超级用户,一个虚拟的上帝。我该怎么办?显然,我不想做任何有破坏性的事情。一位朋友建议为弗雷德·罗杰斯创建一个计算机帐户,他是公共广播系统儿童电视节目“罗杰斯先生的邻居”的传教士和主持人。弗雷德·罗杰斯住在匹兹堡,节目本身是在匹兹堡录制的,距离卡内基梅隆大学仅几步之遥。这个建议看起来非常合适。很快,罗杰斯先生在伯克利机器上获得了他可能的第一个计算机帐户,但他并没有意识到这一点。我们在为罗杰斯先生的帐户命名时犯了一个错误:我的朋友认为弗雷德的中间名首字母是 S,而帐户名是fsr。罗杰斯先生的中间名是麦克菲利。帐户名应该是fmr。

So I became the superuser, a virtual god, on the Berkeley machine. What should I do with the power? Obviously, I did not want to do anything damaging. A friend made the suggestion that a computer account be created for Fred Rogers, the preacher and host of the Public Broadcasting System’s “Mister Rogers’ Neighborhood” TV show for children. Fred Rogers lives in Pittsburgh, and the show itself is recorded in Pittsburgh, within walking distance from Carnegie Mellon. The suggestion seemed very appropriate. Soon Mister Rogers got possibly his first computer account on the Berkeley machine without realizing it. We made one mistake in naming Mister Rogers’ account: my friend thought Fred’s middle initial was S, and the account name was given as fsr. Mister Rogers’ middle name is McFeely. The account name should have been fmr.

伯克利计算机上的一位常规用户显然正在为 DARPA 的一个网络项目工作,他有一个文件,其中包含阿帕网各地计算机的密码。这文件受到保护,但不针对超级用户。密码也是纯文本的。突然间,我有了阿帕网许多常锁前门的钥匙。

One of the regular users on the Berkeley machine was apparently working on a network project for DARPA, and he had a file that contained passwords for machines all over the Arpanet. The file was protected, but not against a superuser. The passwords were also in plain text. Suddenly, I had keys to a lot of normally locked front doors all over the Arpanet.

在接下来的两周里,在我的空闲时间里,罗杰斯先生的邻居开始在阿帕网上扩张。这段时间,我的脑海里不断地响起节目中罗杰斯先生的歌曲:“这是这个社区美好的一天。对于邻居来说是美好的一天。你愿意成为我的吗?你能成为我的吗?……”

For the next two weeks, during my free time, Mister Rogers’ Neighborhood started to expand over the Arpanet. During this period, I was constantly hearing in my mind Mister Rogers’ song from the show, “It is a beautiful day in this neighborhood. A beautiful day for a neighbor. Would you be mine? Could you be mine? …”

为罗杰斯先生创建新帐户进展缓慢,因为每个系统都有不同的弱点。我也没有太多的空闲时间——我所修的三门课程已经接近期末考试了。但在巅峰时期,罗杰斯先生在美国顶尖大学拥有近十个计算机帐户。

Creating new accounts for Mister Rogers was slow going as each system had a different weakness. I also did not have a lot of free time—it was close to the finals for three of the courses that I was taking. But at the peak, Mister Rogers had almost ten computer accounts at top US universities.

然后我收到了我们部门主管尼科·哈伯曼 (Nico Habermann) 的传票。

Then I received a summons from our Department Head, Nico Habermann.

普渡大学的系统经理注意到了为弗雷德·罗杰斯创建的一个帐户,他沿着线索回到了卡内基梅隆大学。后来卡内基梅隆大学的工作人员找到了我。

The system manager at Purdue University had noticed an account created for Fred Rogers and he followed the trail back to Carnegie Mellon. The staff at Carnegie Mellon then found me out.

尼科给了我一个简短的教训,并命令我写下相当于忏悔的内容,说明我是如何做到的,并表达我的悔恨。那是一巴掌打在了手上。没有造成真正的损害,而且相关系统可能因此变得更安全。当然,罗杰斯先生失去了所有账户。这是在国会通过有关计算机隐私的法律之前。如果我的恶行发生在今天,我就会陷入大麻烦。未经许可查看受保护的文件也是不道德的,甚至未经许可也不应查看未受保护的个人文件。我把它当作一个游戏,但事实上这是一件严肃的事情。另一方面,人们永远不应该将包含密码的纯文本文件留在周围,无论是否受保护,并且绝对绝对不要将其保留在具有不需要密码的来宾帐户的系统上。世界很大,也有一些坏人。

Nico gave me a short lecture and I was ordered to write what amounted to a confession about how it was done, and to show my remorse. It was a slap on the hand. There was no real damage, and possibly the systems involved became safer as a result. Mister Rogers, of course, lost all his accounts. This was before Congress passed the law about computer privacy. I would be in deep trouble if my misdeed were done today. It is also unethical to look at protected files without permission, and even unprotected files of a personal nature should not be looked at without permission. I was treating it as a game when in fact it was a serious matter. On the other hand, one should never leave a plain text file containing passwords lying around, protected or not, and definitely, definitely, not on a system with guest accounts that do not require passwords. The world is a big place, and there are some bad people out there.

那一年的黑色星期五会议上,我一定已经危险地接近被指明出路了。事实上,我可能还差一件事情就被要求离开该部门。

I must have been dangerously close to being shown the way out in the Black Friday meeting that year. As it was, I was probably one more incident away from being asked to leave the Department.

    这个错误的假设至今仍然存在。几乎所有有关“深蓝”比赛的国际象棋书籍都错误地指出 Berliner 是 ChipTest/Deep Thought 团队的指导老师。

    This incorrect assumption persists to this day. Almost all the chess books on the Deep Blue matches state wrongly that Berliner was the faculty advisor of the ChipTest/Deep Thought team.

第 4 章中的“来自加利福尼亚州的邀请”    部分详细介绍了整个事件的进展情况。

    The section An Invitation from California in Chapter 4 details how this whole incident played out.

    有关台湾兵役的讨论,请参阅《台湾少年》附录 A中的《避开当局》。

    See Sidestepping Authorities in A Lad in Taiwan, Appendix A, for a discussion on the military service in Taiwan.

    Sharon现任卡内基梅隆大学计算机科学系副系主任和计算机科学学院学术事务助理院长。

    Sharon is now the Associate Department Head of the Computer Science Department and the Assistant Dean for Academic Affairs of the School of Computer Science at Carnegie Mellon.

    操作系统可以被视为控制计算机所有硬件的基本程序。普通程序通过操作系统间接访问系统硬件。

    The operating system can be viewed as the basic program controlling all hardware aspects of a computer. Normal programs indirectly access the system hardware through the operating system.

    工作到很晚的学生往往是黑客。黑客更有可能禁用其终端的写入权限。

    Students working late tended to be hackers. Hackers were more likely to disable the writing privilege to their terminals.

    嗯,不完全是这样。我有朋友在伯克利学习。另一件事是我对计算机游戏《Rogue》中的一些奇特错误感到好奇,该游戏是由伯克利分校的学生编写的。

    Well, not precisely. I had friends studying at Berkeley. Another thing was that I was curious about some of the peculiar bugs in the computer game Rogue, which was authored by students at Berkeley.

 

 

第3章

CHAPTER 3

冒险一试

Taking the Plunge


日本的职业道德

Japanese Work Ethic

我可以选择努力工作,但基本上我是一个懒惰的人。当我在卡内基梅隆大学读二年级的时候,我只努力过一次很长的时间,那就是准备去台湾的高考1。但为了考试而学习并不能真正被视为工作。从这个意义上说,直到我在卡内基梅隆大学的第二年,我一生中从未努力工作过。我的德国同事安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 工作很努力,但他大部分时间都在晚上工作,他的工作态度并没有影响到我。当我开始与一位日本访客合作完成孔教授的一个项目时,情况发生了变化。

I can work hard when I choose to, but basically I am a lazy person. By the time I was in my second year at Carnegie Mellon, I had only worked hard once for an extended period of time, and that was to prepare for my College Entrance Exam in Taiwan1. But studying for exams cannot really be considered work. In that sense, until my second year at Carnegie Mellon, I had never worked hard in my life. Andreas Nowatzyk, my German office mate, was a hard worker, but he worked mostly at night and his work attitude did not rub off on me. This changed when I started to work with a Japanese visitor on one of Professor Kung’s projects.

Teiji Nishizawa是来自日本松下电器的访客。(松下在美国可能更出名的是松下品牌消费电子产品背后的公司。)当时,计算机科学系与日本大公司达成了协议,允许他们在收费的情况下派遣代表临时在教授下设的系。Teiji就是其中的代表之一。

Teiji Nishizawa was a visitor from Matsushita Electric in Japan. (Matsushita is probably better known in the US as the company behind the Panasonic brand of consumer electronic products.) At that time the Computer Science Department had deals with big Japanese companies that allowed them, for a fee, to send representatives to work temporarily at the Department under the professors. Teiji was one of these representatives.

根据我的经验,勤奋的人有两种类型。第一种往往会在短时间内发挥作用。许多软件程序员都属于这种类型,尽管并非所有优秀的程序员都属于这种类型。这种类型的勤奋工作者可能会连续工作 48 小时,同时他们的创造力源源不断。我绝对不能这样工作;如果我睡眠不足,我就会暴躁。Teiji 是第二种勤奋工作者的一个很好的例子。他非常积极主动并且自律。他可能不会工作很长时间,但他会长时间地投入工作。这种勤奋的人更适合芯片设计工作,能够提供持续、高质量的输出比突然闪现的光彩更重要。在接下来的一年左右的时间里,与 Teiji 的密切合作对我来说是一次非常有教育意义的经历,特别是在职业道德方面。

In my experience, there are two types of hard worker. The first tends to work in short bursts. Many software programmers fall into this type, although not all of the good programmers do. Hard workers of this type might work for, say, 48 hours in a row while their creative juices are flowing. I can never work in this way; I get cranky if I don’t have enough sleep. Teiji was a good example of the second type of hard worker. He was highly motivated and self-disciplined. He might not work long hours, but he would put in solid days of work for long periods of time. This type of hard worker is more suitable for chip design work, where being able to provide sustained, high-quality output is more important than sudden flashes of brilliance. Working closely with Teiji for the next year or so turned out to be quite an educational experience for me, especially in terms of work ethic.

Teiji、Alan Sussman(Kung 教授的另一位顾问)和我一起参与了 Kung 教授与通用电气 (GE) 的联合项目。对于艾伦和我来说,这个项目被视为地区资格赛,这是获得博士学位的必备要素。领域资格赛通常是导师给学生的一些任务,以证明他们在研究领域有足够的基础。事实上,这也是教师从学生身上榨取有用作业的一种方式。在这个项目中我第一次接触到了日本的职业道德。

Teiji, Alan Sussman (another of Professor Kung’s advisees), and I worked together on Professor Kung’s joint project with General Electric (GE). For both Alan and me, this project counted as the area qualifier, a required element in getting our PhD degrees. The area qualifier was usually some task given by the faculty advisor to the students to demonstrate adequate grounding in the research area. In reality, it was also a way for the faculty to extract useful work from the students. I got my first contact with the Japanese work ethic during this project.

我们的项目是设计一款芯片,作为多个商用浮点运算芯片2之间的可编程通信通道。可能的应用之一是用于医学诊断的计算机断层扫描,这恰好是通用电气的重要业务。在项目的早期阶段,Kung 和我们三个人每周都会开会来审查设计。当 Teiji 带着手绘的完整顶层原理图以及基本时序信息来参加会议时,我们仍处于指定芯片功能的阶段。由于从未见过有人以这种方式工作,我个人记下了:“哇,这家伙工作得很快。”

Our project was to design a chip to act as the programmable communication channels between several of the commercially-available floating-point arithmetic chips 2. One of the possible applications was computer tomography for medical diagnosis, which happened to be important business for GE. In the early phase of the project, Kung and the three of us had weekly meetings to go over the design. We were still at the stage of specifying what the chip would do when Teiji came to a meeting with hand-drawn, complete top-level schematics, along with the basic timing information. Having never seen anyone working this way, I made a personal note, “Wow, this guy works fast.”

GE负责芯片设计的后端,包括物理布局和芯片制造,而卡内基梅隆大学负责逻辑设计。Teiji、Alan 和我将逻辑设计分为三个部分,我们每个人负责一个部分。一开始,只有 Teiji 投入了时间,但 Alan 和我很快就抓住了精神,并从他的榜样中学习。安德烈亚斯·诺瓦齐克实际上也提供了相当多的帮助。我们正在使用一套新的计算机辅助工程(CAE)工具,Andreas 恰好是第一个用户。

GE was responsible for the back end of the chip design, which included the physical layout and the chip fabrication, while Carnegie Mellon was responsible for the logic design. Teiji, Alan and I split the logic design into three sections, with each one of us responsible for a section. In the beginning, only Teiji was putting in the hours, but both Alan and I soon caught the spirit and learned from his example. Andreas Nowatzyk actually helped quite a bit as well. We were using a new set of Computer Aided Engineering (CAE) tools, of which Andreas happened to be the first user.

联合项目的卡内基梅隆部分于1985年初完成。此时,我已经完成了博士学位的所有其他要求,只剩下提案、研究、写作和论文答辩等小事了。

The Carnegie Mellon portion of the joint project was completed in early 1985. At this point, I had finished all the other requirements for the PhD degree, with only the small matters of proposing, researching, writing, and defending my thesis left to do.

那时,这个联合项目已经持续了一年多。我无法将其用作论文主题,因为没有重要的研究内容。尽管如此,项目经验还是让我变得更加自律,并且,我有生以来第一次经历了芯片的整个逻辑设计过程。如果没有经历这个项目,没有我们的日本访客的微妙影响,我认为我还没有为下一步做好准备。

The joint project had taken over a year by then. I could not use it as a thesis topic, since there was no significant research content. Still, the project experience had made me better disciplined and, for the first time in my life, I had gone through the entire logic design process for a chip. Without going through the project and without the subtle influence of our Japanese visitor, I don’t think I would be ready for my next step.

印象分

Brownie Points

大约在 1984 年,日本办公设备制造商佳能宣布推出一款低成本激光打印机引擎,该引擎有可能用于生产价格低于 3000 美元的激光打印机。一年之内,惠普和苹果都宣布推出首款基于佳能引擎的个人激光打印机,并创建了一个新的细分市场。在佳能宣布之前,激光打印机主要用于大型部门,价格超过 10,000 美元。事实证明,个人激光打印机业务对于惠普和苹果来说都非常重要。对于惠普来说,这一发展是其计算机打印机业务取得巨大成功的开始。对于苹果公司来说,个人激光打印机提供了杀手级应用程序——桌面出版——使得苹果公司的麦金塔电脑能够打入商业市场。

In 1984 or thereabouts, the Japanese office equipment manufacturer Canon announced a low cost laser printer engine that potentially could be used to produce laser printers priced under $3000. Within a year, both Hewlett Packard and Apple announced the first personal laser printers based on the Canon engine, and a new market segment was created. Before Canon’s announcement, laser printers were mostly used in large departments, and cost more than $10,000. The personal laser printer business turned out to be very important for both Hewlett Packard and Apple. For Hewlett Packard, this development was the beginning of their highly successful computer printer business. For Apple, the personal laser printer provided the killer application—desktop publishing—that allowed Apple’s Macintosh computer to penetrate the business market.

随着这些新型激光打印机的普及,能够将电子文档从一个用户发送到另一个用户变得越来越重要。为了与其他用户通信,电子文档的格式必须以某种方式标准化。施乐是激光打印机业务的先驱,并拥有接近现有标准的标准,但该公司在广泛采用其标准方面进展缓慢。Adobe,一个小软件加利福尼亚州的一家初创公司抓住了这个机会,并迅速制定了自己的新页面描述语言Postscript,成为公认的标准。如今,Adobe 的业务蒸蒸日上,市值也超过了施乐公司3

With the proliferation of these new laser printers, it became increasingly important to be able to send electronic documents from one user to another. To communicate with another user, the format of the electronic documents would have to be standardized in some way. Xerox pioneered the laser printer business and had what was close to an existing standard, but the company was slow in getting its standard adopted widely. Adobe, a small software startup in California, seized the opportunity and quickly made its own new page description language, Postscript, the accepted standard. Today, Adobe has a thriving business and a market capitalization greater than Xerox itself3.

我从本科生起就一直关注打印机业务,我发现佳能宣布引发的一系列事件非常有趣。佳能的引擎每分钟最多可以打印八页,但 1985 年使用的打印机控制器在以最大速度打印即使是简单的页面时也存在问题。如果想高速打印复杂的汉字,那就运气不好了。此外,Xerox 和 Adob​​e 的标准都不处理中文字符。有机会生产定制的 VLSI 打印机控制器,可以高速打印更复杂的页面。还有机会创建一个新标准或扩展两个潜在标准之一来处理中文、日语和其他东亚语言。商业机会似乎就在那里,这些想法可能会成为一个不错的博士论文主题。我把我的想法告诉了孔教授,1985年初,我开始收集有关打印机、计算机字体等方面的文献。

I had been following the printer business since I was an undergraduate, and I found the chain of events started by Canon’s announcement quite interesting. Canon’s engine could print up to eight pages per minute, but the printer controller in use in 1985 had problems printing at the maximum speed even for simple pages. If one wanted to print complicated Chinese characters at high speed, one would be out of luck. Furthermore, neither Xerox’s nor Adobe’s standard handled Chinese characters. There was a window of opportunity to produce a custom VLSI printer controller that could print more complicated pages at high speed. There was also an opportunity to either create a new standard or extend one of the two potential standards to handle Chinese, Japanese, and other East Asian languages. The business opportunities seemed to be there and the ideas probably would make a decent PhD thesis topic. I told Professor Kung about my ideas and, in early 1985, I started to collect literature on printers, computer fonts, and so on.

在美国收集相关的中文文献非常困难。然而,我们与GE一起设计的芯片论文在1985年5月在台湾的一个会议上被录用,我有免费的4次回台湾收集中文文献的机会。一切看起来都很好。

It was hard to collect the relevant Chinese literature in the United States. However, a paper on the chip we designed with GE was accepted for a conference in Taiwan in May 1985, and I had a free 4 trip back to Taiwan to collect Chinese literature. Everything was looking good.

然后意想不到的事情发生了。Hans Berliner 博士是一名人工智能教员,他请我帮助设计 Hitech,这是他和一群研究生正在建造的一款国际象棋机器。

Then something unexpected happened. Dr Hans Berliner, who was an AI faculty member, asked me to help with the design of Hitech, a chess machine that he and a group of graduate students were building.

我在卡内基梅隆大学参加的第一次研讨会上就了解了 Hitech 国际象棋走法生成器,它用于直接在 Hitech 硬件中生成国际象棋走法。本次研讨会为VLSI小组研讨会;卡尔·埃伯林是演讲者。走法生成器采用 64 芯片设计,8x8 棋盘的每平方一个芯片。这是很棒的东西,但即使是第一次看了一眼,心里有一个挥之不去的疑问。真的有必要使用64位芯片吗?

I had learned about the Hitech chess-move generator, which was used to generate the chess moves directly in hardware for Hitech, at the very first seminar that I attended at Carnegie Mellon. The seminar was a VLSI group seminar; Carl Ebeling was the speaker. The move generator was a 64-chip design, one chip per square of the 8x8 chessboard. It was neat stuff, but even on first glance, there was a lingering question at the back of my mind. Was it really necessary to use 64 chips?

当 Berliner 博士找到我时,Hitech 的早期版本刚刚设计和制造出来。我并不是 Berliner 接触的第一个有芯片设计经验的研究生。早在我到达卡内基梅隆大学之前,柏林纳就说服卡尔·埃贝林设计了高科技移动发生器。卡尔在同意建造整个国际象棋机器后,后来成为了 Hitech 的首席设计师。

At the time that Dr Berliner approached me, an early version of Hitech had just been designed and built. I was not the first graduate student with chip design experience that Berliner had approached. Well before I arrived at Carnegie Mellon, Berliner had persuaded Carl Ebeling to design the Hitech move generator. Carl later became the chief architect of Hitech after agreeing to build the whole chess machine.

国际象棋机器具有三个主要组件:寻找国际象棋棋步的着法生成器、评估国际象棋机器向前看时所到达位置的质量的评估函数以及指导分析国际象棋机器检查的棋步序列的搜索控制。典型的国际象棋程序具有相同的基本结构,除了着法生成器、评估函数和搜索控制都在软件中。您可以在商店购买的商业“国际象棋机器”也是如此,因为它们实际上只是封装在独立微型计算机中的国际象棋程序。当计算机国际象棋研究人员谈论国际象棋机器时,他们通常谈论的是至少具有硬件移动生成器和硬件评估功能的机器。您可以预期,硬件国际象棋机器的运行速度至少比采用相同半导体技术制造的微处理器上运行的国际象棋程序快一百倍。当 Berliner 博士与我交谈时,Hitech 拥有 64 芯片的移动生成器和非常简单的硬件评估功能。

A chess machine has three main components: the move generator which finds the chess moves, the evaluation function which assesses the quality of the positions reached when the chess machine looks ahead, and the search control which guides analysis of move sequences examined by the chess machine. Typical chess programs have the same basic structure, except that the move generator, the evaluation function, and the search control are all in software. This is also the case for commercial “chess machines” that you can buy in a store, as they are really just chess programs encapsulated in a stand-alone microcomputer. When a computer chess researcher talks about a chess machine, they are usually talking about a machine that has at least a hardware move generator and a hardware evaluation function. You can expect a hardware chess machine to be at least a hundred times faster than chess programs running on microprocessors manufactured in the same semiconductor technology. When Dr Berliner talked to me, Hitech had the 64-chip move generator and a very simple-minded hardware evaluation function.

Berliner 博士希望改进 Hitech 的评估功能。他问我是否有兴趣设计一个 64 芯片系统来计算棋盘上的方格受棋子控制的程度。我告诉 Berliner 博士,我有一个论文题目想要研究,但如果他的问题可以在夏天完成,我可能愿意研究它。我不介意帮柏林博士一个忙,但我不想给他留下这样的印象:他可以抢走我的时间。我不得不承认,我也很期待有机会与卡尔一起工作,他正在组装一个真正的系统,而我从来没有做过这样的事情。当然,整个想法似乎是暂时逃离论文工作的好方法。

Dr Berliner wanted to improve Hitech’s evaluation function. He asked me whether I would be interested in designing a 64-chip system to compute how well the squares of the chessboard were controlled by the chess pieces. I told Dr Berliner that I had a thesis topic that I wanted to work on, but if his problem was something that could be done in a summer, I might be willing to look into it. I did not mind doing Dr Berliner a favor, but I did not want to give him the impression that he could just grab hold of my time. I have to admit that I was also looking forward to the chance of working with Carl who was putting together a real system, and I had never done anything like that. And, of course, the whole idea seemed like a good temporary escape from thesis work.

我与柏林博士会面,了解他的想法。谈话快结束时,他提到我从这项工作中不会得到任何比“布朗尼点”更多的东西。这对我来说是一个新术语。柏林纳解释说,这意味着除了对所做工作的赞赏之外,我不会得到任何有形的东西。我一开始并没有期待任何事情,从我只愿意在夏天做这件事的事实来看,这一点应该很清楚。谈话给我一种奇怪的感觉。

I had a meeting with Dr Berliner to find out what he had in mind. Near the end of the conversation, he mentioned that I would not be getting anything more than “brownie points” from this work. That was a new term for me. Berliner explained that it meant I would not get anything tangible other than appreciation for the work done. I was not expecting anything to begin with, and that should have been quite clear from the fact that I was only willing to work on it for the summer. The conversation gave me an odd feeling.

我在 Berliner 博士的项目上花了大约一周的时间,进入了电路原理图的前几层。然后我与 Hitech 小组举行了一次会议,并介绍了我的一些发现。好消息是“是的,这是可以做到的”。坏消息是“64芯片分区需要添加大量外部电路,系统速度会很慢”。一年后,我发现我对“相当慢”的定义对于 Hitech 来说已经足够快了,但会议时没有人意识到这一点。Berliner 博士似乎很不高兴,并要求我无论如何都要将外部电路安装到芯片上。我解释说,当较慢的片上 FET(场效应晶体管)驱动器取代较快的片外双极驱动器 5 时,系统会变得更慢,因为芯片输入/输出延迟会增加。会议没有做出任何决定就休会了。

I spent about a week on Dr Berliner’s project, entering the top few levels of the circuit schematics. I then had a meeting with the Hitech group and presented some of my findings. The good news was “yes, it could be done”. The bad news was “the 64-chip partitioning makes it necessary to add a lot of external circuits and the system would be quite slow”. A year later, I discovered that my definition of “quite slow” would have been more than adequately fast for Hitech, but nobody realized this at the time of the meeting. Dr Berliner seemed unhappy and asked me to fit the external circuits onto the chip anyway. I explained that the system would get even slower as the chip input/output delays went up when the slower on-chip FET (Field Effect Transistor) drivers replaced the faster off-chip bipolar drivers5. The meeting was adjourned without any decision.

我不是一个快乐的露营者。问题的根源似乎是 64 芯片分区。当我第一次听说 Hitech 项目时,我脑海中浮现的问题又回来了。分区有必要吗?

I was not a happy camper. The root of the problems appeared to be the 64-chip partitioning. The question at the back of my mind when I first heard about the Hitech project came back. Was the partitioning necessary?

20 世纪 80 年代初,来自贝尔实验室的肯·汤普森 (Ken Thompson) 和乔·康登 (Joe Condon) 用中小型集成电路制造了专用国际象棋机器 Belle。1983年,Belle成为第一台参加美国国家大师级别的国际象棋自动机。当我还在台湾服兵役时,它就取代了Chess 4.X成为新的顶级国际象棋自动机。我是到卡内基梅隆大学之后才知道Belle的存在的。

In the early 1980s, Ken Thompson and Joe Condon, both from Bell Laboratories, built Belle, a special purpose chess machine, from small and medium scale integrated circuits. Belle became the first chess automaton to play at the US National Master level in 1983. It had replaced the Chess 4.X as the new top chess automaton while I was still carrying out my military service in Taiwan. I only learned of Belle’s existence after arriving at Carnegie Mellon.

他们写了一篇关于 Belle 的论文,描述了它的基本架构,我在家里重印了一份。那天晚上,与海泰集团开完会后,我拿起重印本重读了一遍。它。是的,Belle论文中描述的着法生成器确实无法安装到单个芯片中。但是否可以修改设计以使其适合呢?

They had written a paper about Belle, describing its basic architecture, and I had a reprint at home. That night, after the meeting with the Hitech group, I picked up the reprint and reread it. Yes, the move generator described in the Belle paper indeed could not be fitted into a single chip. But was it possible to modify the design so that it would fit?

在我看来,有两个主要问题。

The way I saw it, there were two major problems.

首先是美女国际象棋棋步生成器的晶体管太多。电路尺寸太大了。

The first was that the Belle chess move generator had too many transistors. The circuit size was simply too big.

在国际象棋程序中,走法生成器实际上生成程序检查的国际象棋走法。但为了正常工作,走法生成器需要做的不仅仅是生成走法。它还需要由程序生成未经检查或尚未搜索的动作。在 Belle 设计中,第二个任务(仅生成未经检查的动作)是禁用堆栈的工作。Belle 中的禁用堆栈是一个 64 位6宽的存储器,棋盘的每个方格对应一个位。

In a chess program, the move generator literally generates the chess moves that the program examines. But to work properly, the move generator needs to do more than just generate moves. It also needs to only generate unexamined, or not yet searched, moves by the program. In the Belle design, this second task—of generating only unexamined moves—is the job for the disable-stack. The disable-stack in Belle is a 64-bit6 wide memory, one bit for each square of the chessboard.

禁用堆栈对于搜索的每个动作都有两个条目。正常的国际象棋棋步有起始方(攻击者)和终止方(受害者)。禁用堆栈上的每个条目对应于两种类型的方块之一。

The disable-stack has two entries for every move searched. A normal chess move has a from-square (attacker) and a to-square (victim). Each entry on the disable-stack corresponds to one of the two types of squares.

Belle 中的移动生成分为两个阶段。在第一阶段(寻找受害者),会生成一个新的 to-square(受害者)。例如,第一阶段可能会生成“rook on h8”作为受害者。禁用堆栈上的第一个条目用于屏蔽已为此阶段搜索的所有 to-squares。如果“rook on a8”是已搜索到的受害者,则 a8 方块将被屏蔽。

Move generation in Belle has two phases. In the first phase (find-victim), a new to-square (victim) is generated. For instance, the first phase might generate “rook on h8” as the victim. The first entry on the disable-stack is used to mask off all the to-squares already searched for this phase. If “rook on a8” is a victim already searched, then the a8 square is masked off.

找到受害者后,贝儿进入移动生成的第二阶段。在第二阶段(寻找攻击者)中,为给定的受害者找到新的方(或攻击者)。假设我们可能会发现“knight on g6”是“rook on h8”的攻击者。禁用堆栈上的第二个条目用于屏蔽已搜索给定受害者的所有来自方(攻击者)。如果没有剩下攻击者,那么贝尔将返回第一阶段寻找新的受害者。

After a victim has been found, Belle enters the second phase of move generation. In this second phase (find-attacker), a new from-square (or attacker) is found for the given victim. Say we might find that “knight on g6” is an attacker for “rook on h8”. The second entry on the disable-stack is used to mask off all the from-squares (attackers) that are already searched for the given victim. If no attacker is left, then Belle returns to the first phase to find new victims.

在我后来设计的所有国际象棋机器中,我允许程序向前查找 128 层7 。每次移动两个条目,禁用堆栈需要 256 个字深才能处理多达 128 层的前瞻层。让我们假设每一位内存都需要六个晶体管。单独禁用堆栈所使用的晶体管数量对于棋盘的每个方格来说至少是 1500 个,或者对于整个棋盘来说大约是 100,000 个晶体管。以今天的标准来看并不多。但早在 1985 年,在添加了所有其他电路后,将 Belle move 发生器安装到单个芯片中几乎是不可能的。

In all the chess machines that I later designed, I allowed the program to look up to 128 plies7 ahead. With two entries per move, the disable-stack would need to be 256 words deep to handle up to 128 plies of looking ahead. Let us assume that we need six transistors for every bit of memory. The number of transistors used by the disable-stack alone would be at least 1500 for each square of the chessboard, or about 100,000 transistors for the whole board. Not a lot by today’s standards. But back in 1985, after adding in all the other circuits, it was nearly impossible to fit the Belle move generator into a single chip.

有什么可能的方法来减少晶体管的数量?我的公寓里已经有《百丽报》的重印两年多了,但我从未思考过这个问题,原因有二。一是我不相信计算机国际象棋问题的解决方案是可以实现的。因此,我没有动力去进一步探究。第二是我做出了错误的假设。我认为 Belle 的设计师以及 Hitech 团队一定已经探索了减少晶体管数量的方法。我忘记了 Belle 的设计师虽然很有能力,但并没有试图将他们的设计安装到单个芯片上。他们用分立的现成芯片设计了 Belle,而用于禁用堆栈的芯片相当容易获得。对于他们来说,禁用堆栈可能是最优雅的解决方案。它只是不是单芯片设计的最佳解决方案。事实上,Hitech 集团的国际象棋走棋生成器采用了 64 芯片设计,这也影响了我的判断。如果 Hitech 团队需要 64 个芯片来完成他们的走棋生成器,那么有人创建单芯片国际象棋走棋生成器的可能性有多大?

What possible ways were there to reduce the number of transistors? I had had the reprint of the Belle paper in my apartment for more than two years, but had never thought about this question for two reasons. One was that I did not believe that a solution to the Computer Chess Problem was within reach. Therefore, I had no motivation to probe further. The second was that I had made a wrong assumption. I assumed that the Belle designers, as well as the Hitech group, must have explored ways to reduce the number of transistors. I forgot that the Belle designers, able though they were, were not trying to fit their design onto a single chip. They designed Belle with discrete off-the-shelf chips, and chips used for the disable-stack were fairly easy to come by. For them, the disable-stack was perhaps the most elegant solution. It just was not the best solution for a single chip design. The fact that the Hitech group had adopted the 64-chip design for their chess move generator also clouded my judgment. If the Hitech group needed 64 chips to complete their move generator, how likely was it for anyone to create a single chip chess move generator?

我躺在床上,第一次认真思考禁用堆栈的问题。当我开始思考这个问题时,我意识到可能确实有一个解决方案。Belle 的设计师并不是在寻找替代解决方案,而 Berliner 博士可能很高兴 Hitech 的设计与 Belle 的设计不同。当 Carl Ebeling 接手创建 Hitech 走法生成器时,Hitech 团队就已经确定了使用 64 个筹码的想法,即棋盘的每平方一个筹码。因此,卡尔集中精力完善64芯片设计。因此,尽管所有相关人员能力都很强,但他们没有充分的理由去寻找替代方案。很可能有一个禁用堆栈的替代方案!

There I was, lying on my bed, thinking about the question of the disable-stack seriously for the first time. Once I started thinking about the question, I realized there might indeed be a solution. The Belle designers were not looking for an alternative solution and Dr Berliner was probably quite happy that the Hitech design was different from the Belle design. When Carl Ebeling took the job of creating the Hitech move generator, the Hitech group was already set with the idea of using 64 chips, one chip per square of the chessboard. Consequently, Carl concentrated all his efforts on perfecting the 64-chip design. So, even though all the people involved were highly capable, they did not have strong reasons to look for alternatives. There may very well be an alternative to the disable-stack!

最根本的答案就是完全摆脱移动生成器芯片的禁用堆栈。最简单的方法是从移动生成器中取出禁用堆栈芯片并连接 64 根电线,从移动生成器芯片到外部禁用堆栈,棋盘的每个方格一根电线。这是一个可能的解决方案,但我不喜欢它。理想情况下,我更喜欢将任何芯片的8引脚数量保持在40 以下。当时,更高引脚数的封装可能花费超过 25 美元。按照这个价格,封装的成本可能比芯片本身还要高。想要将引脚数控制在 40 个以下的一个更个人的原因是看看我能将这个想法推进到什么程度。Hitech move 发电机芯片采用 40 引脚封装。如果不仅可以实现单芯片设计,而且新设计可以使用与 Hitech 芯片相同的封装,那将是一个非常妙的主意。我不喜欢外部禁用堆栈解决方案的另一个原因是它实际上是九芯片设计。64 位宽的外部禁用堆栈需要八个内存芯片。当时最宽的存储芯片只有8位宽。

The most radical answer would be simply to get rid of the disable-stack from the move generator chip completely. The easiest way to do this was to take the disable-stack out of a move generator chip and connect 64 wires, one wire for each square of the chessboard, from the move generator chip to the external disable-stack. It was a possible solution but I did not like it. Ideally, I preferred to keep the number of pins8 for any chip below 40. Higher pin-count packages back in those days could cost over twenty-five dollars. At that price, the package might cost more than the chip itself. A more personal reason for wanting to keep the pin count under forty was to see how far I could push the idea. The Hitech move generator chips were using 40-pin packages. It would be quite a coup if not only a one-chip design was possible, but also the new design could use the same package used by the Hitech chips. Another reason why I did not like the external disable-stack solution was that it was really a nine-chip design. Eight memory chips would be needed for the 64-bit wide external disable-stack. The widest memory chip in those days was only 8-bit wide.

这时,我想到了一个想法。新想法消除了禁用堆栈的整个设计,而不仅仅是芯片本身。

At this point, an idea hit me. The new idea rid the whole design of the disable-stack, not just the chip itself.

这个想法取决于禁用堆栈真正携带的信息。禁用堆栈正在跟踪迄今为止搜索的移动的历史记录。在国际象棋机器中,“移动堆栈”包含完全相同的信息,尽管形式不同。理论上,存储在禁用堆栈中的 64 位模式可以根据上次搜索的移动并当前存储在移动堆栈上来计算。最后一个移动块的起始方和终止方包含计算 64 位禁用掩码所需的所有信息。我坐起来,抓起笔和纸,开始计算细节。棋盘的六十四个方格中的每一个都被分配了一个唯一的优先级编号。对于给定的起始方或终止方,我们禁用具有较高优先级的所有方格。我们可以使用 6 到 64 优先级解码器9电路从 6 位平方数计算 64 位禁用掩码。优先级解码器电路使用了大约十个晶体管每个方块的平均值,而不是禁用堆栈使用的 1500 个晶体管。这是 150 比 1 的减少。这一减少幅度有多大?我最终实现的完整移动生成器平均每个方格使用了大约五百五十个晶体管。如果不改变设计,每平方的晶体管数量将超过 2000 个,芯片尺寸将增大四倍。

The idea hinged on what information the disable-stack was really carrying. The disable-stack was keeping track of the history of the moves searched so far. In a chess machine, the “move stack” contains the very same information, albeit in a different form. The 64-bit patterns stored in the disable-stack could, in theory, be computed from the last move searched and currently stored on the move stack. The from-square and the to-square of the last moving piece contain all the information needed to compute the 64-bit disable-mask. I sat up, grabbed a pen and paper, and started working out the details. Every one of the sixty-four squares of the chessboard is assigned a unique priority number. For a given from-square or to-square, we disable all squares with higher square priorities. We can compute the 64-bit disable-mask from the 6-bit square number by using a six to sixty-four priority decoder9 circuit. The priority decoder circuit used about ten transistors on average for each square instead of the fifteen-hundred transistors used by the disable-stack. This was a 150 to 1 reduction. How significant was this reduction? My final implementation of the complete move generator used about five-hundred and fifty transistors on average for each square. Without the design change, the number would have been over two-thousand transistors per square, and the chip would have been too large by a factor of four.

我高兴得目瞪口呆。但贝尔的设计还有第二个主要问题。这个问题可能会让外行人感到惊讶。对于年轻得多的我来说,这肯定会感到惊讶。当我还是个小孩子的时候,更高质量的收音机是带有更多真空管或更多晶体管的收音机。对于硅芯片来说,还有什么比其晶体管数量更重要的呢?答案是电线的数量,特别是长电线的数量。Belle move 发电机的电线太多。

I was happily stunned. But there was a second major problem with the Belle design. This problem may come as a surprise to a layperson. It certainly would have come as a surprise to a much younger me. When I was a little kid, higher quality radios were the ones with more vacuum tubes or more transistors. What could be more important to a silicon chip than the number of transistors it has? The answer is the number of wires, in particular, the number of long wires. The Belle move generator had too many long wires.

Belle 移动生成器为棋盘的 64 个方格中的每一个生成 6 个棋子优先级信号,每种棋子类型一个。在基于禁用堆栈按棋子位置进行简单移动排序之前,棋子优先级信号用于按棋子的值对移动进行排序。一般来说,我们应该首先搜索以最不有价值的攻击者俘获最有价值的受害者的着法。这种基于价值的走棋顺序可以显着提高国际象棋机器的效率。

The Belle move generator produces six piece-priority signals, one per piece type, for each of the sixty-four squares of the chessboard. The piece-priority signals are used to order the moves by the values of the pieces, before the simple move ordering by piece locations based on the disable-stack. In general, we should firstly search moves that capture the most valuable victim with the least valuable attacker. This value-based move ordering can improve the efficiency of the chess machine dramatically.

384 (6x64) 棋子优先级信号被送入优先级编码器,这是一个用于选择最佳着法的投票电路。问题出在 384 个片优先信号上。它们是 384 根长电线。

The 384 (6x64) piece-priority signals get fed into a priority encoder, which is a voting circuit used to select the best move. The problem lies with the 384 piece-priority signals. They are 384 long wires.

对于可用的 3 微米 CMOS 技术,仅 384 条长电线就可以覆盖芯片的一半。我需要一种替代方式来进行动议投票。我花了不到一分钟就找到了可行的替代方案。这完全是运气问题。那时候,我订阅了很多IEEE(电气和电子工程师协会)的期刊。(IEEE 为学生订阅期刊提供了大幅折扣。) 在那个决定命运的夜晚之前几周,我在IEEE 计算机杂志上读到了一篇关于名为“未来总线”的新计算机总线标准的文章。未来总线使用“分布式仲裁器”电路在总线上的电路板之间进行投票。“分布式仲裁器”需要很少的电线,可能很适合解决我在 Belle move 发电机中使用长电线的问题。这本杂志就放在我旁边。快速检查该文章证实,稍加修改的分布式仲裁员会完成这项工作。长电线的数量不再是 384 根,而是减少到 48 根,减少了八分之一。

For the 3-micron CMOS technology that was available, the 384 long wires alone would have covered half of the chip. I needed an alternative way to do the move voting. It took me less than one minute to find a working alternative. It was all a matter of luck. Back in those days, I subscribed to a lot of IEEE (Institute of Electrical and Electronic Engineers) journals. (IEEE gave students hefty discounts for journal subscriptions.) A few weeks before that fateful night, I had read an article in IEEE Computer Magazine about a new computer bus standard called Future Bus. The Future Bus used a “distributed arbiter” circuit to vote among circuit boards residing on the bus. The “distributed arbiter” required few wires and might be a good fit for my problem with the long wires in the Belle move generator. The magazine was lying right next to me. A quick check of the article verified that a slightly modified distributed arbiter would do the job. Instead of 384 long wires, the number of long wires would be reduced to forty-eight, a reduction by a factor of eight.

因此,很有可能仅使用 40 引脚封装将改进的 Belle move 发生器安装到单个芯片上。事实上,我们有理由相信,这样的棋子生成器也能比Hitech使用的64芯片棋子生成器快十倍以上。最终真正的芯片速度接近二十倍。其成本效益高出一千倍 (64 × 20)。

So, it might very well be possible to fit a modified Belle move generator onto a single chip using only a 40-pin package. As a matter of fact, there were reasons to believe that such a move generator could be more than ten times faster than the 64-chip move generator used in Hitech as well. The real chip in the end was closer to a factor of twenty times faster. It was over a thousand times (64 × 20) more cost effective.

事情突然变得非常有趣。到目前为止,我一直认为计算机象棋问题是遥不可及的。64芯片Hitech移动发生器的设计似乎证实了我的信念。如果需要六十四个VLSI芯片才能完成一个相对较慢的移动发生器,那么这个问题确实是遥不可及的。但现在看来你可以只用一个芯片构建一个移动生成器。计算机国际象棋问题值得重新审视。

Things were suddenly very interesting. Up to this point, I had always assumed that the Computer Chess Problem was out of reach. My belief seemed to have been confirmed by the 64-chip Hitech move generator design. If it took sixty-four VLSI chips to complete a relatively slow move generator, then the problem was indeed beyond reach. But now it appeared that you could build a move generator with only one chip. The Computer Chess Problem deserved a second look.

在 Belle 国际象棋机器中,除了着法生成器之外,评估函数是机器其余部分的主要部分。评估功能也可以安装在单个芯片上吗?一个好的评估函数将比移动生成器复杂得多。如果几乎不可能将移动生成器安装到单个芯片上,那么为获得良好的评估函数而做同样的事情的机会有多大?看起来不太好,是吗?只有一个选择:即,我们可以用空间换取时间。国际象棋评估函数具有空间重复成分,并且可以多次使用较小的电路来进行相同的计算,尽管会带来时间损失。一个有趣的事实是,我们不必一直付出时间代价。评估函数可以分为两个部分:材料(基本上是留在棋盘上的棋子的价值)和位置(任何其他东西)。可以非常快速地计算材料评估。位置评估是耗时的部分。通常,位置评估也具有相对较小的值,通常小于棋子的值。如果物质平衡相差太远,比如一侧是皇后,则可以忽略位置评估函数并且不需要计算。后来在真实国际象棋机器上测量的统计数据表明,仅在大约 10-20% 的时间内需要进行全面评估。及时复用较小的电路以减少芯片面积似乎是一个切实可行的选项。粗略计算表明,64 芯片评估函数没有意义,此时,我失去了帮助 Berliner 博士使用他的 64 芯片评估函数的兴趣。Hitech 系统的速度足够慢,因此时分复用单芯片评估功能仍然足够快。

In the Belle chess machine, besides the move generator, the evaluation function was the bulk of the rest of the machine. Could the evaluation function be fitted onto a single chip as well? A good evaluation function will be far more complicated than the move generator. If it was barely possible to fit the move generator onto a single chip, what was the chance of doing the same for a good evaluation function? It doesn’t look good, does it? There was one option: namely, we could trade space for time. Chess evaluation functions have spatially repetitive components, and it is possible to use a smaller circuit multiple times to do the same computation, albeit with a time penalty. An interesting fact is that we do not have to pay the time penalty all the time. The evaluation function can be partitioned into two components: material (basically the values of the pieces left on the board) and positional (anything else). The material evaluation can be computed very fast. The positional evaluation is the part that is time consuming. Usually, the positional evaluation also has relatively small values, frequently smaller than the value of a pawn. If the material balance is too far off, say one side is down a queen, the positional evaluation function can be ignored and does not need to be computed. Statistics measured later on real chess machines indicated that the full evaluation is only needed about 10–20% of the time. Multiplexing a smaller circuit in time to reduce the chip area appeared to be a viable option. A back-of-an-envelope calculation showed me that the 64-chip evaluation function did not make sense and at this point, I lost interest in helping Dr Berliner with his 64-chip evaluation function. The Hitech system was sufficiently slow that a time-multiplexed single chip evaluation function would still be more than fast enough.

第二天,我给 Berliner 博士发了一封电子邮件,告诉他我不再对构建 64 芯片评估函数感兴趣,并且可以构建单芯片国际象棋着法生成器和单芯片评估函数。如果他有兴趣的话,我还提出做一个简单版本的单芯片评估功能。Berliner 博士的第一反应是试图让我“以正确的方式设计整个芯片”。也就是按照他指定的方式来设计芯片,采用64芯片的设计。我明确表示我不想做64芯片的设计。经过一周的反复讨论,我向 Berliner 博士发送了一份简单时分复用电路的一页初步规范,该电路计算了评估函数的子集。柏林纳博士对我的提议并不完全满意。我决定坚持自己的立场,并表示如果我的建议不被接受,我就不想参与。也许我有点太直率了。此时我不打算在计算机国际象棋方面做更多的事情,而且我有一个严肃的论文提案需要完成。

The next day I sent an e-mail to Dr Berliner telling him that I was no longer interested in building the 64-chip evaluation function, and that it was possible to build single chip chess move generators and single chip evaluation functions. I also offered to do a simple version of the single chip evaluation function, if he was interested. Dr Berliner’s first reaction was to try to get me to “design the whole chip the right way”. That is, to design the chip as he specified, using the 64-chip design. I made it clear that I didn’t want to do the 64-chip design. After a week of going back and forth, I sent Dr Berliner a one-page preliminary spec of a simple time multiplexed circuit that calculated a subset of the evaluation function. Dr Berliner was not entirely satisfied with what I was proposing. I decided to stand my ground and say that I did not want to be involved if my proposal was not acceptable. Perhaps I was a little too blunt. I was not planning to do anything more with computer chess at this point, and I had a serious thesis proposal to work on.

不久之后,我见到了孔教授,并向他讲述了发生的事情。孔教授强烈建议我写下单芯片设计的想法并向几位教员做演示。孔教授没有明说,但从他的语气中,我感觉我可能会在黑五教员会议(每学期末召开的教员会议,决定每一个研究生的命运)上遇到大问题。部门)。在我对 Berliner 博士的直率回应中,我实际上声称 Hitech 已经过时了,但没有提出理由。由于我之前在阿帕网的冒险行为,我在教员眼中的地位已经摇摇欲坠。如果我写论文并做演示,那么教授们就会知道我的说法并非轻率。在接下来的几周里,我放弃了论文提案的工作,并写了一份技术报告。在我飞往台湾的飞机前一天,我还向教职员工做了一次演讲。

Shortly after, I had a meeting with Professor Kung and told him about what had happened. Professor Kung strongly suggested that I write up the ideas for the single chip designs and give a presentation to a few faculty members. Professor Kung did not say it, but from his tone, I got the feeling that I might have major problems at the faculty Black Friday meeting (the faculty meeting held at the end of each semester to decide the fate of each and every graduate student in the department). In my blunt response to Dr Berliner I had effectively claimed that Hitech was passé, without presenting the reasoning. My standing in the eyes of the faculty was already shaky from my earlier escapade over the Arpanet. If I wrote the paper and did the presentation, then the faculty would know that I did not make the claim frivolously. For the next few weeks, I dropped my work on my thesis proposal and wrote up a technical report. I also gave a presentation to faculty members the day before I was to get on the plane to Taiwan.

1985 年 5 月的大部分时间我都在台湾度过。部分时间是为了打印机控制器的想法,花了很多时间收集有关汉字、书法等的书籍。剩下的时间大部分都是在度假。我当时正在和自己争论是否要进入电脑象棋领域。我去卡内基梅隆大学不是为了从事电脑游戏工作;从事计算机国际象棋工作会与我最初的兴趣有很大的偏差。此外,这一承诺很可能需要五到十年的时间。然而,撰写技术报告和进行演示让我意识到,我已经有了构建“国际象棋机器之母”的基本蓝图,这是一台可以击败世界冠军的机器。换句话说,我有机会追求计算机科学中最古老的圣杯之一,并可能创造历史。从消极的一面来看,打印机控制器看起来也很有前途,而且这不是一个可以等待的想法。个人激光打印机仍处于起步阶段,在行业标准制定之前,定制打印机控制器芯片被广泛采用的时间窗口有限。打印机控制器的想法有可能使我在财务上独立。解决计算机国际象棋问题几乎不可能提供与所需时间和努力相称的经济回报。如果我要解决计算机象棋问题,我会因为知道我解决了这个问题而感到荣耀。我可能不会在经济上陷入贫困,但会有重大的个人牺牲。机会成本将非常高。我可能把一生中最富有成效的岁月都花在一个回报微薄的项目上。

I spent most of May 1985 in Taiwan. Part of the time was spent collecting books about Chinese fonts, calligraphy and so on for the printer controller idea. The rest of the time was mostly just vacationing. I was debating with myself whether I wanted to go into the computer chess arena. I did not go to Carnegie Mellon to work on computer games; working on computer chess would be quite a deviation from my original interest. Moreover, the commitment could easily take five to ten years. Writing the technical report and doing the presentation, however, made me realize that I had the basic blueprint to build the Mother of all Chess Machines, a machine that could defeat the World Champion. In other words, I had a chance to pursue one of the oldest holy grails in computer science, and possibly make history. On the negative side, the printer controller looked quite promising as well, and it was not an idea that could wait. Personal laser printers were still in their infancy and there was a limited time window to make a custom printer controller chip widely adopted before the industry standard set in. There was some chance that the printer controller idea could make me financially independent. There was practically no chance that solving the Computer Chess Problem would ever provide any financial reward commensurate with the time and efforts required. If I went after the Computer Chess Problem, I would be going for the glory of knowing that I solved the problem. I probably would not be financially destitute, but there would be major personal sacrifices. The opportunity cost would be extremely high. I could be spending many of the most productive years of my life on a project that had a very slim payoff.

这是一个艰难的决定。最终,经过认真的自我反省,我决定去追求荣耀。你不可能每天都有机会创造历史。

The decision was a difficult one. In the end, after serious soul searching, I decided to go for the glory. You don’t get a chance to make history every day.

至少两年

At Least Two Years

有时,没有什么比在正确的时间出现在正确的地点更重要的了。

Sometimes, nothing is more important than being at the right place at the right time.

如果我没有在卡内基梅隆大学,我就不会有迈克·布朗和安德烈亚斯·诺瓦齐克作为我的办公室伙伴。如果迈克和安德烈亚斯对国际象棋或电脑国际象棋没有兴趣,那么当柏林博士请我帮忙时,我对电脑国际象棋的兴趣可能已经完全消失了。如果在那个决定命运的夜晚我的公寓里没有《Belle》论文和《IEEE》杂志,我可能就不会想出如何安装国际象棋机到单芯片上。如果柏林纳博士接受了我的建议,我可能会尽我所能并继续我最初的研究项目。如果我在其他计算机科学系,如果孔教授不是我的指导老师,我可能不会有作为研究生的自由来自己启动一个项目。我在 1985 年开始这个项目的事实也很重要。五年前,该技术还不足以制造单芯片国际象棋走棋发生器。五年后,你将能够将 Belle move 生成​​器按原样安装到单个芯片上,而其他人会比我先一步做到这一点。我走上这条路或多或少是一个意外。

If I had not been at Carnegie Mellon, I would not have had Mike Browne and Andreas Nowatzyk as my office mates. If Mike and Andreas had no interest in chess or computer chess, my interest in computer chess probably would have completely dissipated by the time Dr Berliner asked me to help. If I had not had both the Belle paper and the IEEE magazine in my apartment on that fateful night, I probably would not have figured out how to fit a chess machine onto a single chip. If Dr Berliner had accepted my suggestion, I probably would have just done what I could and gone on with my original research project. If I had been at some other Computer Science Department, and, if Professor Kung had not been my faculty advisor, I might not have had the freedom as a graduate student to start a project on my own. The fact that it was 1985 when I began the project was also important. Five years earlier, the technology would not have been sufficiently advanced for fabricating a single chip chess move generator. Five years later you would be able to put the Belle move generator, as it was, onto a single chip, and somebody else would have beaten me to it. It was more or less an accident that I ended up on the path that I took.


当我第一次成为孔教授的顾问时,比“麻烦制造者”的标签早很多年,他告诉我,研究生才是系里真正有权力的人,因为他们才是真正掌握研究项目的人。孔教授可能是鼓励我要更加自信,因为很多亚洲学生都有不发言的问题。

When I first became Professor Kung’s advisee, many years before the troublemaker label, he told me that graduate students were the ones with the real power in the Department, as they were the ones with a true grasp of the research projects. Professor Kung was probably encouraging me to be more assertive, as many Asian students had the problem of not speaking up.

从台湾回来后,我告诉孔教授我决定研究计算机象棋问题。孔教授很惊讶,但也接受了。他暗示我可能只有有限的时间来研究芯片。他没有提及原因,但无需天才就能意识到我不能花太长时间。该系已经有一个计算机国际象棋项目。如果没有快速的进展,开始一个新的项目将是一种学术自杀。

After my return from Taiwan, I told Professor Kung about my decision to go after the Computer Chess Problem. Professor Kung was surprised but receptive. He hinted that I might have only limited time to work on the chip. He did not mention the reason, but it did not take a genius to realize that I could not afford to take too long. The Department already had a computer chess project. Starting a new one without fast progress would be an academic suicide.

在我能够全职研究国际象棋芯片之前,发生了另一件事。我回来几天后,Kung 教授告诉我,GE 的人在模拟我们与他们一起设计的芯片时遇到了问题,他们需要帮助。

Before I could work full time on the chess chip, something else happened. A few days after my return, Professor Kung informed me that the GE people were having problems simulating the chip that we designed with them and they needed help.

此时 Teiji 已经返回日本,艾伦·萨斯曼和我是唯一能够提供帮助的人。我们俩都想花时间进行自己的研究,但我们一开始就没有报名参加模拟。最终,我们俩都参与其中。我最终接手了大部分工作,因为我意识到这并不是浪费时间,至少对我来说不是。我们与 GE 一起设计的芯片具有相当多的晶体管级电路,因此我们必须进行大量的“开关级”模拟,其中晶体管被模拟为开关。新的移动生成器尚未设计还没有,但我知道它主要是动态10晶体管级电路,GE 芯片的经验将非常有用。

By this time Teiji had already returned to Japan, and Alan Sussman and I were the only people who might be able to help. Both of us wanted to spend time on our own research, and we did not sign up to do the simulation to begin with. Eventually, we both got involved. I took over the bulk of the work in the end, as I realized that it was not a waste of time, at least not in my case. The chip we designed with GE had quite a bit of transistor level circuitry, so we had to do a significant amount of “switch level” simulation, where the transistors were simulated as if they were on-off switches. The new move generator had not been designed yet, but I knew that it would be mostly dynamic10 transistor level circuitry, and the experience from the GE chip would be quite useful.

GE的工作直到1985年8月下旬才完成。就在完成之前,我接手了VLSI CAD(计算机辅助设计)工具的维护,因为之前的维护者刚刚成为父亲。接管这项工作符合我自己的利益,因为我将在一段时间内成为这些工具的唯一用户。我已经准备好开始我的新项目了。

The GE work was not finished until late August 1985. Just before its completion, I took over the maintenance of the VLSI CAD (Computer Aided Design) tools, as the previous maintainer had just become a father. It was in my own interest to take over the work, as I was to become the only user of the tools for a while. I was ready to begin my new project.

Carl Ebeling 设计的 Hitech 64 芯片国际象棋走法生成器花了三年多的时间才完成。我没有那么多时间花在我的着法生成器芯片上。我计划在五到十年内完成击败世界冠军的目标,但我不能只花三年时间在移动生成器上。我决定全力以赴,在六个月内完成芯片。为什么是六个月的目标?好吧,乔·康登 (Joe Condon) 和肯·汤普森 (Ken Thompson) 在六个月内完成了 Belle 国际象棋机器,如果他们能够在这段时间内构建出一个完整的国际象棋机器,我应该能够完成一个 VLSI 国际象棋移动生成器。当时,卡内基梅隆大学以及其他所有大学的所有重要芯片项目都花了一年多的时间才能完成。如果我成功了,这将是一个速度记录。

The Hitech 64-chip chess move generator designed by Carl Ebeling took over three years to complete. I did not have that much time to spend on my move generator chip. I planned to complete my goal of beating the World Champion in five to ten years, and I could not afford to spend three years on just the move generator. I decided to go full steam ahead and complete the chip in six months. Why the six-month target? Well, Joe Condon and Ken Thompson finished the Belle chess machine in six months, and if they could build a complete chess machine in that time, I should be able to complete a VLSI chess move generator. At the time, all the chip projects of any significance at Carnegie Mellon, and possibly every other university, took over a year to complete. If I pulled it off, it would be a speed record of sorts.

该项目的第一阶段,即电路设计,非常简单,并且很大程度上借鉴了我在 GE 联合项目中的经验。虽然很辛苦,但一路上并没有什么大的惊喜。移动发生器的完整电路设计,一直到晶体管级别,在一个月内完成。同一时期还对整个芯片进行了整体模拟,直至晶体管开关级别。

The first phase of the project, the circuit design, was straightforward, and largely drew from my experience with the GE joint project. It was hard work, but there was no major surprise along the way. The complete circuit design of the move generator, down to the transistor level, was finished in one month. The full chip was also simulated in its entirety down to the transistor switch level in the same period.

接下来是困难的部分。物理布局——绘制所有几何图形,包括晶体管和电线——是下一步。如今,大多数设计师都会使用自动化工具来进行布局。1985 年可用的工具无法做到这一点,所以很大程度上我必须手动进行布局。但还有另一个问题。我想将芯片装入 40 引脚封装中,但此类芯片的芯片尺寸受到严重限制。目前量产芯片的最大芯片尺寸约为 2 厘米(0.79 英寸)在一边。MOSIS 11 40 引脚封装中可容纳的最大芯片的一侧边缘约为 0.7 厘米(0.28 英寸),大约相当于当今大型芯片面积的 12%。我的任务是将设计融入有限的模具区域,并使用手头的工具来完成。还有一个小问题,我对芯片设计的物理布局步骤基本上没有经验。

Next came the hard part. The physical layout—the drawing of all the geometry, including the transistors and the wires—was the next step. Nowadays, most designers would be using automated tools to do the layout. The tools available in 1985 could not do it, so to a great extent I had to do the layout manually. But there was another problem. I wanted to fit the chip into a 40-pin package and the die size for such chips was severely limited. The largest die size for chips in production quantities today are around 2 cm (0.79 in) on a side. The maximum die that could fit into the MOSIS11 40-pin package was about 0.7 cm (0.28 in) on one edge, or roughly twelve percent of the area of today’s big chips. My task was to fit the design into the limited die area and to do it with the tools at hand. There was also the small problem that I had essentially no experience with the physical layout step of the chip design.

在电路设计阶段,我将棋盘 64 个方格中每个方格的核心逻辑减少到平均约 550 个晶体管。由于边缘效应以及棋子在不同等级上以不同方式移动的事实,方块的电路并不相同,但它们确实共享许多共同的逻辑。第一步,我选择了一个正方形并对该正方形进行了尝试布局。这花了大约一个月的时间。试验布局完成后,有好消息也有坏消息。好消息是,根据试验布局所使用的面积,该芯片可能适合大量工作。坏消息是我为一个正方形设计的试验布局形状错误。它太高了。为了获得最大的芯片面积,本质上,芯片必须是正方形的。允许的最大尺寸为 6812 x 6912 微米,或 0.6812 厘米 x 0.6912 厘米。由于棋盘的尺寸为 8 × 8,因此每个逻辑方格的电路也必须以方格的形式进行物理布局。

In the circuit design phase, I reduced the core logic to about 550 transistors on average for each of the sixty-four squares of the chessboard. Owing to edge effects and the fact that a pawn moves in a different way on different ranks, the circuits for the squares were not identical, but they did share a lot of common logic. As a first step, I picked one of the squares and did a trial layout of the square. This took about a month. At the completion of the trial layout, there was good news and bad news. The good news was, that based on the area used by the trial layout, the chip might fit with a lot of work. The bad news was that the trial layout I had for one square had the wrong shape. It was too tall. To get the maximum chip area, essentially, the chip would have to be a square. The maximum dimensions allowed were 6812 by 6912 microns, or 0.6812 cm by 0.6912 cm. Since a chessboard is 8 × 8, the circuit for each logical square would have to be physically laid out as a square as well.

接下来的两个月用于纠正布局的形状并为所有不同的正方形创建变化。当为了减少芯片面积,需要将方块平铺到 X 和 Y 维度上精确配合时,不同方块的数量增加了四倍。通过将正方形平铺至精确配合,我可以在相邻正方形之间共享电线和触点(IC 上电线之间的连接),并为每对相邻正方形挤出几微米。进行平铺的最简单方法是在 X 和 Y 方向上镜像正方形。这种镜像意味着,即使对于具有相同电路的正方形,也有四种不同版本的布局:东北、西北、东南部和西南部。芯片核心的最终布局,逻辑上是一个 8 × 8 棋盘,结果看起来更像是一个 4 × 4 图案。图案相当漂亮,有点像被子。(当基于该芯片的国际象棋机 Deep Thought 广为人知时,一位海报供应商问我他是否有权根据骰子照片制作海报。)虽然最终结果可能在美学上令人愉悦,但两者几个月的辛苦工作对我个人来说简直就是地狱。这项工作很乏味,花费的时间比我预期的要多得多。我必须工作到深夜的时间越来越长才能跟上我的日程安排。当然,在逻辑设计阶段,我的周末已经提前变成了工作日。

The next two months were spent correcting the shape of the layout and creating variations for all the different squares. The number of different squares quadrupled when, in order to reduce the chip area, it became necessary to tile the squares to an exact fit in both X and Y dimensions. By tiling the squares to an exact fit, I could share wires and contacts (connections between wires on an IC) between adjacent squares and squeeze out a few microns for every pair of adjacent squares. The easiest way to do the tiling was to mirror the squares in both X and Y directions. This mirroring meant that, even for squares with identical circuits, there were four different versions of the layout: northeastern, northwestern, southeastern, and southwestern. The final layout of the chip core, which was logically an 8 × 8 chessboard, looked more like a 4 × 4 pattern as a result. The pattern was rather pretty and somewhat quilt-like. (When Deep Thought, the chess machine based on this chip, became widely known, a poster vendor asked me whether he could have the right to produce posters based on the die photo.) While the end result might have been aesthetically pleasing, the two months of hard work had been hellish for me personally. The work was tedious and consumed far more time than I had expected. I had to work longer and longer into the night to keep up with my schedule. Of course, my weekends had already become workdays earlier during the logic design phase.

此时已是十二月,又一个黑色星期五即将到来。孔教授让我给高科技小组做一个演讲。Berliner 博士在演讲前开始表现出一些兴趣,但我们的兴趣再次并不一致。我只是想建造一台非常快的机器,看看它能做什么,然后决定下一步做什么。柏林纳博士有一个完全不同的想法。

It was December by this time, and there was another Black Friday coming up. Professor Kung asked me to give a presentation to the Hitech group. Dr Berliner started to show some interest before the presentation but, again, our interests did not really align. I was simply going to build a very fast machine, see what it could do, and decide what to do next. Dr Berliner had a completely different idea.

柏林纳博士是计算机国际象棋领域的权威,他的话很有分量。他公开表示,像我计划的那样,强力12搜索不会走得太远,是时候回到选择性搜索13的旧想法了,特别是 B 星的想法(B *) 搜索算法,这是 Berliner 博士在 20 世纪 70 年代提出的算法。大多数计算机国际象棋研究人员似乎都同意 Berliner 博士的观点,即暴力搜索已达到其极限,但 B* 可以统治世界的建议并未得到普遍接受。

Dr Berliner was a leading authority in the computer chess field, and his words carried a lot of weight. He was publicly stating that brute force12 searching, such as I was planning, would not go very far and it was time to go back to the old idea of selective searching13, in particular, to the idea of the B-star (B*) search algorithm, an algorithm that Dr Berliner had come up with in the 1970s. Most of computer chess researchers appeared to agree with Dr Berliner that brute force searching was reaching its limit, but the suggestion that B* could rule the day was not universally accepted.

我的直觉是 B* 是一条死胡同。不管怎样,我做了演讲,Berliner 博士给了我关于 B* 的参考资料。我没有确实有时间仔细检查他的参考资料,直到芯片完成。

My intuition was that B* was a dead end. Anyway, I gave my presentation and Dr Berliner gave me references about B*. I did not really have the time to go over his references until the chip was done.

安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 与 Hitech 集团的几位成员经常接触。在我演讲几天后,他告诉我,Hitech 团队,或者至少是团队中的某个人,相信我需要“至少两年”才能完成芯片设计。我真是被逗乐了。我刚刚向 Hitech 小组介绍了一款完全模拟的芯片,几乎完全布局,但至少有人仍然相信我还需要两年时间才能完成设计。我已经完成了所有方块的布局,我可以看到我很有机会按照我的计划在六个月内完成芯片。两年的预测可能是基于卡尔完成高科技移动发生器所需的时间,大约是三年。我不一定是比卡尔更好的芯片设计师,但我有更好的工具,我不必向不时更改设计规范的委员会负责,而且我决心在尽可能短的时间内完成芯片可能的。如果这意味着我必须在这个项目上持续花费很长时间,我会毫不犹豫地这样做。除了所有这些原因之外,我还得到了卡尔的间接帮助。在研究 Hitech move Generator 的过程中,他制作了一种工具,可用于将一个网表(所有晶体管的列表以及晶体管之间的连接)与另一个网表进行比较。通过将布局的网表与逻辑设计的网表进行比较,设计人员可以验证布局是否与设计匹配。Carl 的工具在设计的最终验证阶段非常重要。

Andreas Nowatzyk had frequent contact with several members of the Hitech group. A few days after my presentation, he told me that the Hitech group, or at least someone in the group, believed that it would take me “at least two years” to finish the chip design. I was quite amused. I had just given the Hitech group a presentation about a chip that was completely simulated, and almost completely laid out, and yet at least one person still believed that it would take me two more years to complete the design. I had completed the layout of all the squares, and I could see that I had a good chance to finish the chip in six months as I planned. The two-year prediction was probably based on how long it had taken Carl to finish the Hitech move generator, which was about three years. I was not necessarily a better chip designer than Carl, but I had better tools, I did not have to answer to a committee who changed the design spec from time to time, and I was determined to finish the chip in as short a time as possible. If it meant that I had to spend long hours on the project for sustained periods, I would not hesitate to do so. On top of all these reasons, I had Carl’s indirect help. During his work on the Hitech move generator, he produced a tool that could be used to compare one netlist (the list of all the transistors and the connections between the transistors) with another. By comparing the netlist for the layout with the netlist for the logic design, a designer can verify that the layout matches the design. Carl’s tool was quite important in the final verification phase of the design.

经过两个月的时间布置棋盘上的所有方格后,我遇到了一个新问题要处理。即使所有的方块都紧密地结合在一起,芯片的边缘也没有留下太多的空间。在剩下的空间里,我不得不挤进方块的边缘布线、芯片的控制逻辑、芯片外部世界的接口逻辑、接口逻辑和 IO(输入/输出)焊盘之间的布线,当然,还有 IO 焊盘本身。主要的问题是垫子。MOSIS 提供的现有垫根本不适合。焊盘通常通过机械接合的金线将芯片连接到外界。鉴于引线键合工艺的机械性质,与其他芯片功能相比,IO 焊盘的引线着陆区域相当大。然而,接线区域仅是 IO 焊盘的一小部分。剩下的该区域被特殊的保护电路和IO驱动电路所占据,保护电路保护芯片免受外部静电放电的影响,IO驱动电路将内部信号增强到足够的强度以驱动外部电路。MOSIS IO 焊盘是长矩形,导线着陆区域靠近矩形的短边之一。这些焊盘的问题在于焊盘的长边垂直于芯片边缘。如果您需要在芯片周围放置一百多个 IO 连接,那么它们就很好,但如果您不需要大量 IO 连接,则不太好。使用 MOSIS 焊盘,芯片周围的 IO 环将是一个非常厚的环。如果我决定使用 MOSIS 焊盘,则移动生成器芯片的宽度和高度大约会多出 1000 微米,或者就芯片面积而言大约会多出 30%。

After two months of laying out all the squares of the chessboard, I had a new problem to deal with. Even with all the squares fitting tightly together, there was not much space left on the rim of the chip. In the remaining space, I had to squeeze in the edge wiring for the squares, the control logic for the chip, the interface logic to the world outside the chip, the wiring between the interface logic and the IO (input/output) pads, and, of course, the IO pads themselves. The main trouble was with the pads. The existing pads available from MOSIS simply would not fit. The pads connect the chip, usually through mechanically bonded gold wires, to the outside world. Given the mechanical nature of the wire bonding process, the wire landing area for an IO pad is fairly large in comparison to other chip features. The wire landing area, however, is only a small portion of an IO pad. The rest of the area is taken up with special protection circuits, which protect the chip from external electrostatic discharges, and the IO driver circuit, which boosts the internal signal to adequate strength for driving external circuits. The MOSIS IO pads were long rectangles, with the wire landing area near one of the short edges of the rectangle. The problem with these pads was that the long edges of the pads were perpendicular to the chip edges. They were fine if you needed to place, say, more than a hundred IO connections around the chip, but not so good if you didn’t need a lot of IO connections. With the MOSIS pads, the IO ring around the chip would be a very thick ring. If I decided to go with the MOSIS pads, the move generator chip would have been about 1000 microns too wide, and too tall, or roughly thirty percent too large in terms of die area.

我设计的移动生成器使用不超过 40 个引脚,因此我确实有一个替代方案,即获取一组新的打击垫。从芯片边缘看,MOSIS 焊盘又高又瘦。我的新护垫必须又短又粗。由于没有具有 IO pad 设计经验的专家,所以我就自己设计了。焊盘和其余芯片电路的布局大约花了一个月的时间。从一开始五个月后,我就完成了完整芯片布局的第一次切割。它几乎不适合允许的最大芯片面积。我在 X 或 Y 方向上的剩余空间不超过 5 微米,甚至不足以再挤一根电线。谈论将其关闭。

I had designed my move generator to use no more than 40 pins, so I did have an alternative, namely, getting a new set of pads. Viewed from the chip edge, the MOSIS pads were tall and skinny. My new pads would have to be short and fat. Since there was no expert with design experience for IO pads available, I just went ahead and designed my own. The layout of the pads and the remainder of the chip circuits took about a month. Five months from the start, I had a first cut of the complete chip layout. It barely fitted into the maximum die area allowed. I had no more than five microns to spare in either X or Y direction—not even enough to squeeze in one more wire. Talk about cutting it close.

此时我感觉良好并且变得大胆。又过了一年,我才向整个系提交了我的论文提案,但我很确定我的论文会是什么。我对外部论文委员会成员的选择也有一个很好的想法。在卡内基梅隆大学,论文委员会通常由三名来自大学本身的成员和一名外部成员组成。孔教授将自动担任论文委员会主席,通常他会在寻找外部成员方面发挥主导作用。我没有告诉孔教授,直接给肯·汤普森发了一封电子邮件,询问他是否介意加入我的论文委员会。我没有告诉肯我还没有论文委员会的小细节。通常,您不仅会邀请图灵奖获得者加入您的委员会;还会邀请图灵奖得主加入您的委员会。我依靠这个请求的震撼力来引起肯的兴趣。他感到相当震惊,并询问有关我工作的更多信息。我把之前写的技术报告发给他了我去台湾旅行,他读完后同意了。我不知道为什么我没有邀请在计算机国际象棋领域非常有名的柏林博士加入我的委员会。可能我只是想强调这篇论文实际上是关于硬件设计和计算理论,而不是计算机国际象棋本身。然而,事后看来,我可能潜意识里也担心潜在的利益冲突。

At this point I was feeling good and getting bold. It would be another year before I presented my thesis proposal to the Department at large, but I was quite certain what my thesis would be. I also had a good idea about my choice of the outside thesis committee member. At Carnegie Mellon, the thesis committee normally had three members from the university itself and one outside member. Professor Kung would automatically be the head of the thesis committee and normally he would play a leading role in finding the outside member. Without telling Professor Kung, I sent an e-mail directly to Ken Thompson, asking whether he would mind being on my thesis committee. I did not tell Ken the minor detail that I did not have a thesis committee yet. Normally, you don’t just ask a Turing Award winner to be on your committee; I was relying on the shock value of the request to get Ken interested. He was suitably shocked, and asked for more information on my work. I sent him the technical report that I had written before my trip to Taiwan and, after reading it, he agreed. I don’t know why I did not ask Dr Berliner, who was very well known in the computer chess field, to be on my committee as well. It might have been that I simply wanted to emphasize that the thesis was really about hardware design and computation theory instead of computer chess per se. With hindsight, however, I might also have had some subconscious concern about a potential conflict of interest.

我原来的六个月计划还剩一个月。该芯片仍需验证。首先,我需要重新仿真以验证芯片的功能。这很简单,但在重新模拟之前我必须做一件事。在布局过程中,我对电路进行了更改,以实现更有效的布局,但我尚未更新电路原理图以反映这些更改。进行了更新并重新模拟了电路。下一步,也就是 Carl 的工具发挥作用的地方,是验证重新模拟的网表和物理网表是否相同,晶体管对晶体管,电线对电线。最后一步有一个小问题——卡尔的工具对于网表的外观有自己的想法。我编写了一个翻译器来创建正确形式的网表,然后比较两个新的网表。通常,大多数设计人员会以自下而上的方式比较网表,即首先比较最小子电路的网表,然后比较更高级别电路的网表,最后比较顶层电路的网表。为了以防万一我很幸运,我首先比较了网表的顶层。

I had one month left of my original six-month schedule. The chip still had to be verified. First, I needed to re-simulate to verify the chip’s functionality. This was straightforward, but there was one thing that I had to do before the re-simulation. During the layout process, I made changes to the circuit to allow more efficient layout, but I had not yet updated the circuit schematics to reflect the changes. The updates were made and the circuit was re-simulated. The next step, and this was where Carl’s tool came in, was to verify that the re-simulated netlist and the physical netlist were identical, transistor for transistor and wire for wire. There was a slight problem associated with this last step—Carl’s tool had its own idea on how the netlists should look. I wrote a translator to create netlists of the right form, and then compared the two new netlists. Usually, most designers would compare the netlists in a bottom-up fashion, that is, starting by comparing the netlists for the smallest sub-circuits, then the netlists for higher levels of circuits, and eventually the netlists for the top level. Just in case I got lucky, I compared the top level of the netlists first.

毫不奇怪,比较失败了。不匹配的地方太多了。特别是,物理网表中电源和地被短路(直接连接)。如果我按原样提交芯片,我就会得到一个可以用来煮鸡蛋的热芯片。在电路的顶层,有数以万计的电源和接地引脚连接。当然,从网表的顶层无法判断电源和接地短路的位置。明显的策略是分而治之,即以自下而上的方式比较两个网表。一旦真正的网表比较开始,我发现了许多其他错误。幸运的是,所有错误都可以在不增加已经很紧的芯片尺寸的情况下得到纠正。到 1986 年 2 月,也就是我开始工作六个月后,芯片已经设计、布局、验证并准备好制造。

Not surprisingly, the comparison failed. There were too many mismatches. In particular, the power and the ground were shorted (directly connected) in the physical netlist. I would have one HOT chip that could be used to boil an egg if I submitted the chip the way it was. On the top level of the circuits, there were tens of thousands of connections to the power and the ground pins. Naturally, it was not possible to tell from the top level of the netlists where the power and the ground were shorted. The obvious strategy was to divide and conquer, namely, to compare the two netlists in the bottom-up fashion. Once the real netlist comparison started, I found numerous other errors. Luckily, all the errors could be corrected without increasing the already tight die size. By February 1986, six months after I started, the chip was designed, laid out, verified and ready for fabrication.

现在,大问题来了。芯片第一次就能工作吗?或者 Hitech 小组中的人预测我至少需要两年才能让我的移动生成器工作是正确的吗?

Now, the big question. Would the chip work the first time? Or would the person in the Hitech group be correct in predicting that it would take me at least two years to get my move generator working?

硅棋盘。

A Silicon Chessboard.

硅画布上的铝,1985-1986 年绘制,徐凤雄。

Aluminum on silicon canvas, drawn in 1985–1986, Feng-hsiung Hsu.

还充当单芯片国际象棋走棋生成器,是 1987 年至 1995 年间顶级国际象棋机器的核心

Also acts a single chip chess move generator and was the core of the top chess machines from 1987 to 1995

汤姆·索亚的诡计

Tom Sawyer’s Trick

半导体行业在 20 世纪 80 年代中期经历了重大转型。20 世纪 80 年代初,NMOS(N 型 MOS [金属氧化物半导体])是低成本、大批量的技术。到 20 世纪 80 年代中期,CMOS(互补 MOS)逐渐取代 NMOS,成为占主导地位的低成本技术。CMOS技术同时使用P型MOS晶体管和N型MOS晶体管,其中P型MOS晶体管的主要载流子具有正电荷,而N型MOS晶体管的主要载流子具有负电荷。CMOS电路功耗较低,但工艺也比NMOS复杂。

The semiconductor industry went through a major transition in the mid-1980s. In the early 1980s, NMOS (N-type MOS [Metal-Oxide-Semiconductor]) was the low cost, high volume technology. By the mid-1980s, CMOS (Complementary MOS) was slowly replacing NMOS as the dominant low cost technology. The CMOS technology uses both the P-type MOS transistor, where the main charge carrier has a positive charge, and the N-type MOS transistor, where the main charge carrier has a negative charge. CMOS circuitry has lower power consumption, but the process is also more complicated than NMOS.

MOSIS,一家提供芯片代工服务的硅经纪商美国大学大约在1985年开始提供CMOS制造服务。一开始,很少有大学研究人员使用CMOS服务。原因一是缺乏CMOS设计经验,二是MOSIS提供的CMOS工艺还不成熟,传闻有人从MOSIS拿到了可以工作的CMOS芯片,但没有人见过。公平地说,提交给 MOSIS 的大多数设计,无论是 NMOS 还是 CMOS,一开始就可能存在缺陷。

MOSIS, the silicon broker providing chip foundry service to US universities, started to offer a CMOS fabrication service in about 1985. In the beginning, few university researchers used the CMOS service. One reason was the lack of experience with CMOS design, and a second was that the CMOS process provided by MOSIS was not mature yet—rumor had it that someone got working CMOS chips from MOSIS, but no one saw them. To be fair to MOSIS, most of the designs submitted to MOSIS, either NMOS or CMOS, were probably flawed to begin with.

我在1986年3月提交了CMOS国际象棋芯片。通常情况下,封装好的芯片大约需要六到八周的时间才能回来,我还有时间做其他事情。我仔细查看了与 B* 相关的参考文献,但没有留下深刻的印象。阅读证实了我的第一直觉14。与说法相反,至少对我来说,这些数字似乎表明 B* 不如简单地搜索每一步棋到相同深度的方法(即所谓的蛮力方法

I submitted the CMOS chess chip in March 1986. Normally, it took about six to eight weeks before the packaged chips came back, and I had some time to work on something else. I took a closer look at the references related to B*, but I was not impressed. The reading confirmed my first intuition14. Contrary to the claims, the numbers appeared, at least to me, to show B* to be inferior to the simple-minded approach of just searching every move to the same depth—the so-called brute force approach.

当然,与柏林博士的某种合作仍然是可能的。毕竟,我的团队需要一名优秀的国际象棋棋手,而柏林纳博士是最好的棋手,他是前世界函授国际象棋冠军。

Of course, some sort of cooperation with Dr Berliner might still be possible. After all, I would need a good chess player on the team, and Dr Berliner was the best player around, being a former World Correspondence Chess Champion.

到目前为止,我假设我会找到某种方法来并行化搜索。我的计划有两个步骤,即创造一台比 Belle 快一千倍的国际象棋机器。第一个是创建一个由单个芯片或少量芯片组成的国际象棋机器。第二个是找到一种方法,可以并行使用大量的棋机来有效地加快搜索过程。鉴于单芯片国际象棋走棋生成器已经设计完成,我已经完成了第一步。然而,第二步则是完全不同的事情。当时,尽管进行了多年的积极研究,但没有人知道如何有效地并行搜索。我只是相信一定会找到解决办法。对我来说,我是否是第一个找到解决方案的人并不重要;重要的是我是否会成为第一个找到解决方案的人。只要在最终与世界冠军相遇之前找到一个就可以了。当然,如果我能自己找到解决方案,那就太好了。

Up to this point, I was making the assumption that I would find some way to parallelize the search. There were two steps in my plan to create a chess machine a thousand times faster than Belle. The first was to create a chess machine that was either a single chip or a small number of chips. The second was to find a way to use a massive number of chess machines in parallel to effectively speed up the search process. Given that the single-chip chess move generator had already been designed, I was part way through the first step. The second step was, however, an entirely different matter. At the time, no one knew how to effectively parallelize the search, despite many years of active research. I simply believed that a solution would be found. It did not matter to me whether I would be the first person to find a solution; as long as one was found before the final encounter with the World Champion, I would be fine. Of course, it would be so much sweeter if I could find the solution myself.

问题比我想象的要简单。到 1986 年 4 月,我有了一个不错的解决方案,初步看来效果很好模拟并可以在理论上证明其具有一些相当好的特性。大约在这个时候,我听说我提交的芯片的制造运行遇到了一些问题,并且需要相当长的时间才能到达芯片。听到这个消息,我很震惊,5月初再次提交了芯片。提交两天后,MOSIS 给我发了一条消息,说我第一次提交的芯片太大,无法装入我指定的 40 引脚封装中。我立即告诉他们芯片的最大尺寸完全正确。几个小时后,我收到了道歉——软件故障在他们的计算中使我的芯片增加了几微米,并且芯片撞到了更大的芯片上。这些芯片将采用 64 引脚封装,大约是 40 引脚封装的三倍。我不确定第二次运行是否会正确进行,考虑到连续的故障,我在五月下旬和六月下旬又提交了两次芯片。

The problem turned out to be simpler than I thought. By April 1986, I had a decent solution that appeared to work well in preliminary simulations and could be proven theoretically to have some fairly good properties. At about this time, I heard that the fabrication run for my submitted chip had run into some problems and it would take quite a while before the chips arrived. Shocked by the news, I submitted the chip again in early May. Two days after my submission, MOSIS sent me a message saying that the chip I submitted in the first run was too large to fit into the 40-pin packages that I specified. I informed them immediately that the chip was at exactly the right maximum size. A few hours later, I got an apology—a software glitch had added a few microns to my chip in their calculation, and the chip got bumped into a larger die. The chips would be packaged in 64-pin packages, which were about three times bigger than the 40-pin packages. I did not know for sure that the second run would go correctly and, given the successive glitches, I submitted the chip twice more in late May and late June.

我于 1986 年 6 月 28 日收到了第一批 13 个芯片。令我惊讶的是,它们是第二批的,采用 40 引脚封装。这次运行使用的铸造厂名声不好,我有点担心。我测量了电源和地之间的电阻,有几个芯片短路了。我的心微微沉了下去。我用来设计芯片的 CAE 工作站有一个附加的硬件连接,可用于测试芯片的功能。我把没有短路的芯片接上了。不去。这些芯片都没有表现出任何生命力——它们甚至没有通过最简单的测试。我打开了一些芯片封装,并将它们放在显微镜下;缺陷是显而易见的。我向 MOSIS 发送了一份报告,告诉他们这些缺陷。他们的第一反应是要求提供其中一块劣质芯片。然后我收到另一个回复,询问我是否参与过任何大型项目或有任何测试芯片的经验。至少可以说,我有点生气。显然,没有其他人在收到芯片的当天报告收益率为零。很可能当时只有我一个人在抱怨。

I received the first thirteen chips on 28 June 1986. To my surprise, they were from the second run and were packaged in 40-pin packages. The foundry used on this run had a bad reputation, and I was a little bit apprehensive. I measured the resistance between power and ground, and several of the chips were shorted. My heart sank a little. The CAE workstation that I used to design the chip had an attached hardware hookup that could be used to test the chips functionally. I hooked up the chips that were not shorted. No go. None of the chips showed any glimmer of life—they were not passing even the simplest test. I opened some of the chip packages and placed the chips under a microscope; defects were plain to see. I sent a report to MOSIS, telling them about the defects. Their first response was a request for one of the bad chips. Then I got another response questioning whether I had worked on any large project or had any experience testing chips. I was more than a bit annoyed, to say the least. Apparently, no other people were reporting zero yields on the very day they received the chips. It was very likely that I was the only one complaining at the time.

两周后,下一批芯片于 7 月 14 日抵达。这些芯片来自期待已久的首轮测试,采用 64 引脚封装。测试设置适用于 40 引脚封装。Andreas 将 40 针插座和 64 针插座焊接在一起作为适配器,我能够在旧设置上使用适配器测试新芯片。这次芯片的表现要好得多。大多数芯片通过了一些测试,其中两个芯片通过了所有测试测试。答对了。该芯片“第一次”工作了。合格芯片的百分比(也称为良率)有点低,并且在显微镜下观察未工作的芯片表明金属线比应有的宽。

The next batch of chips arrived two weeks later on July 14. These were from the long-awaited first run, and were packaged in 64-pin packages. The test setup was for 40-pin packages. Andreas soldered together a 40-pin socket with a 64-pin socket to act as an adaptor, and I was able to test the new chips using the adaptor on the old setup. The chips behaved much better this time. Most of the chips passed some tests, and two of the chips passed all the tests. Bingo. The chip worked the “first time”. The percentage of good chips (also known as the yield) was a little bit low, and looking at the non-working chips under a microscope showed that the metal wires were wider than they should be.

我向 MOSIS 发送了另一份报告,提到这些芯片与零产量运行中的设计相同。这次我收到回复说芯片供应商干法蚀刻(不使用湿法化学品蚀刻)金属层,但MOSIS已经准备了湿法蚀刻的掩模。这解释了本次运行的初始延迟。几天后,第三批产品到达,芯片采用 40 引脚封装。这次运行的产量更高,大约一半的芯片可以工作。第四批产品在一个月后到达,由与第二批产品相同的供应商制造,并且产量再次为零。MOSIS这次认真对待了该报告,将供应商列入观察名单,甚至建议使用国际象棋芯片作为该供应商资格的测试芯片。

I sent out another report to MOSIS, mentioning that these chips were the same design as in the zero-yield run. This time I got a reply indicating that the chip vendor dry etched (etching without using wet chemicals) the metal layer, but MOSIS had prepared the mask for wet etch. This explained the initial delay for this run. A few days later, the third run arrived, with chips in 40-pin packages. This run had better yield and about half of the chips worked. The fourth run arrived a month later, was fabricated by the same vendor as the second run, and again had a zero yield. MOSIS this time took the report seriously, put the vendor on the watch list, and even suggested using the chess chip as a test chip for qualifying the vendor.

当我拿到前两个“工作”芯片时,我验证了这些芯片是否按设计工作。但该设计本身尚未被证实可以在真正的国际象棋程序运行。

When I got the first two “working” chips, I verified that the chips worked as designed. But the design itself was not verified to work within a real chess program yet.

在卡内基梅隆大学,计算机科学系有一个称为工程实验室的支持组织,负责管理系内网络和机器,并为研究人员提供工程支持。那里的几位工程师都是我的好朋友。其中一位劳伦斯·布彻(Lawrence Butcher)借给我一张计算机总线接口卡,并给了我一块空的小绕线板。(绕线板是一种原型板,其中电路连接是通过将电线缠绕在板背面的柱子上来实现的。)我在绕线板上绕线了一个简单的九芯片国际象棋芯片接口。劳伦斯帮我制作了一条连接两块板的电缆。现在的问题是找到一台机器来放置总线接口卡。一种可能性是将其放置在工程实验室的一台机器中,但该实验室与我的办公室位于不同的楼层。然后我想到了一个好主意。

At Carnegie Mellon, the CS Department had a support organization known as the Engineering Lab to take care of the departmental network and machines, and to provide engineering support for the researchers. Several of the staff engineers there were good friends of mine. One of them, Lawrence Butcher, lent me a computer bus interface card and gave me a small empty wire-wrap board. (A wire-wrap board is a prototype board where circuit connections are made by wrapping wires around posts on the backside of the board.) I wire-wrapped a simple nine-chip chess chip interface on the wire-wrap board. Lawrence helped me make a cable connecting the two boards. Now, the problem was finding a machine to put the bus interface card in. One possibility was to put it in one of the machines in the Engineering Lab, but the Lab was on a different floor from my office. Then I hit upon a great idea.

在马克·吐温的经典小说《汤姆·索亚历险记》中,汤姆奉命粉刷栅栏。他假装粉刷栅栏很有趣,让其他孩子觉得这很有趣,所以他们愿意付钱给汤姆为他做工作。作为一个懒惰的人,当我还是个小男孩的时候第一次读到这个故事时,这个故事给我幼小的心灵留下了深刻的印象。

In Mark Twain’s classic novel The Adventures of Tom Sawyer, Tom was ordered to paint the fence. He pretended that it was fun painting the fence, and got the other kids to think that it was such fun that they were willing to pay Tom to do his work for him. Being a lazy person myself, this story made quite an impression on my young mind when I first read it as a small boy.

我的同事迈克·布朗(Mike Browne)在我研究芯片时编写了一个国际象棋程序。我试图引起他对移植他的程序的兴趣与国际象棋芯片一起运行,并不是说我想要有一个有竞争力的国际象棋程序,但我确实想确保国际象棋芯片能够正常工作。如果我能让芯片在一个有效的国际象棋程序中运行,最好是由其他人来完成这项工作,我可以肯定。我与迈克的直接接触并没有奏效。想起《汤姆·索亚历险记》,我采取了一种更偷偷摸摸的方法。

Mike Browne, my office mate, wrote a chess program while I was working on the chip. I tried to interest him in porting his program to run with the chess chip, not that I wanted to have a competitive chess program, but I did want to make sure that the chess chip worked. If I could get the chip to run inside a working chess program, preferably with someone else doing the work, I would be sure. My direct approach with Mike did not work. Remembering Tom Sawyer, I went for a sneakier approach.

Thomas Anantharaman 和我是同一届计算机科学研究生。他在演讲组,但他也在我所在的VLSI组,所以我们有很多接触。托马斯的成长经历很不寻常。他的父亲是一位退休的冶金工程教授,也是一位瑜伽专家。他的德国母亲是印度哲学的学生。托马斯毕业于印度贝拿勒斯印度教大学理工学院。我不知道为什么,但他似乎能够以与我不同的方式思考;有时,他会得出我一千年都做梦也想不到的结论。托马斯的办公室和我的办公室在同一层,而且只有几扇门的距离。

Thomas Anantharaman and I were in the same incoming class of CS graduate students. He was in the speech group, but he was also in the VLSI group that I was in, so we had a lot of contact. Thomas had quite an unusual upbringing. His father, a retired professor of Metallurgical Engineering, was also an expert Yoga practitioner. His German mother was a student of Indian Philosophy. Thomas graduated from the Institute of Technology at Benares Hindu University in India. I don’t know how, but he seems to be capable of thinking in different ways than I do; sometimes, he reaches conclusions that I would never dream about in a thousand years. Thomas’ office was on the same floor as mine and it was only a few doors away.

和迈克一样,托马斯也为了自己的娱乐而编写了一个国际象棋程序,他有一个与我的国际象棋芯片接口兼容的办公室工作站。如果我知道我的下一步行动将如何影响托马斯的论文研究,我就会避免这样做(尽管,归根结底,他最终做得很好)。不管怎样,我说服了托马斯让我把接口卡放到他的工作站上。然后我写了一个更详细的测试程序来测试芯片。每当新一批芯片进来时,我都可以在五分钟内测试它们,每个芯片大约十秒。我将测试程序的源代码放在 Thomas 可以访问的公共场所,然后我回到模拟如何并行化搜索,暂时假装忽略 Thomas。

Like Mike, Thomas had also written a chess program for his own amusement, and he had an office workstation that was compatible with my interface to the chess chip. Had I known how my next action would impact on Thomas’ thesis research, I would have refrained from doing it (although, when all was said and done, he did all right in the end). Anyway, I persuaded Thomas to let me put the interface card into his workstation. I then wrote a more detailed testing program to test the chips. Whenever a new batch of chips came in, I could test them all within five minutes, at about ten seconds per chip. I placed the source of the testing program in a public place that Thomas could access, and then I got back to my simulation on how to parallelize the search, pretending to ignore Thomas for the moment.

几天后,托马斯告诉我,他已经调整了他的国际象棋程序以使用国际象棋芯片,正如我秘密计划的那样。托马斯能够使用国际象棋芯片让他的程序每秒搜索大约 30,000 个位置。如今最快的基于 PC 的国际象棋程序的搜索速度至少快 10 倍,但在 1986 年,只有专用国际象棋机和在超级计算机上运行的国际象棋程序才能更快。托马斯的新程序进行的第一场比赛是对阵默里·坎贝尔(Murray Campbell),他是研究生和国际象棋专家,也是柏林博士的学生。默里在高科技方面做了一些工作,尽管不是他的一部分主要研究。他对于如何与计算机下国际象棋当然有很好的想法。然而,穆雷还是输掉了比赛。经过我自己的详细测试后,我非常确定该芯片是正确的,但看到它在工作的国际象棋程序中使用,消除了任何残余的疑虑。

A few days later, Thomas informed me that he had adapted his chess program to use the chess chip, just as I had secretly planned. Thomas was able to get his program, with the chess chip, to search about 30,000 positions/sec. The fastest PC-based chess programs today can search at least 10 times faster, but in 1986 only special purpose chess machines, and chess programs running on supercomputers, were faster. The first game that Thomas’ new program played was against Murray Campbell, a fellow graduate student and chess expert, as well as a student of Dr Berliner’s. Murray had done some work on Hitech, although not as part of his main research. He certainly had good ideas about how to play chess against computers. Yet, Murray lost the game. I was quite certain that the chip was correct after my own detailed testing, but seeing it used in a working chess program dispelled any residual doubt.

当托马斯参与其中时,我并没有真正计划组建我们自己的计算机国际象棋团队。我原以为我会找到一些共同点并与 Hitech 团队合作。让托马斯移植他的国际象棋程序只是避免做我应该自己做的工作的捷径。我当时的计划是制造一台三片国际象棋机器。除了已经完成的走棋生成器芯片之外,这两个新芯片将评估棋局(评估芯片)并控制搜索(搜索控制器)。由于技术困难,我从未完成芯片组。其中之一是评估芯片上需要有 RAM(随机存取存储器)。在没有详细过程控制的情况下设计 RAM(如 MOSIS 服务的情况)并非易事。另一个主要的技术困难是我当时对国际象棋了解不多。后来当我和托马斯决定参加电脑国际象棋比赛时,芯片组的计划就被永久搁置了。

At the time that Thomas got involved, I had not really planned on forming our own computer chess team. I had assumed that I would find some common ground and work together with the Hitech group. Getting Thomas to port his chess program was just a short cut to avoid doing a piece of work that I should be doing myself. My plan at the time was to create a three-chip chess machine. Besides the already completed move generator chip, the two new chips would evaluate chess positions (the evaluation chip) and control the search (the search controller). I never did complete the chip set because of technical difficulties. One of them was that the evaluation chip would need to have RAMs (Random Access Memory) on it. Designing RAMs without detailed process control, as in the case of MOSIS service, was no trivial task. The other major technical difficulty was that I really did not know that much about chess at the time. When Thomas and I decided to enter into computer chess competitions later on, the plan for the chip set became permanently shelved.

    有关台湾教育制度的讨论,请参阅《台湾小伙子》中的早起者,附录 A。

    For a discussion of the Taiwanese educational systems, see An Early Riser in A Lad from Taiwan, Appendix A.

    计算机中使用两种类型的算术:定点和浮点。定点运算最简单的例子是整数运算,其中固定二进制小数点位于整数的最低有效位(二进制数字)之后。浮点运算使用两个数字(大小尾数)来表示单个数量。幅度部分保存有效位;尾数是指示二进制小数点所在位置的数字,可以将其视为比例因子。浮点运算通常用于工程应用。20 世纪 80 年代初,在单芯片上独立进行浮点运算成为可能。如今,浮点运算单元仅是通用微处理器的一小部分。

    There are two types of arithmetic used in computers: fixed point and floating point. The simplest example of fixed-point arithmetic is integer arithmetic, where the fixed binary point is after the least significant bit (binary digit) of the integer. Floating point arithmetic uses two numbers, magnitude and mantissa, to represent a single quantity. The magnitude part holds the significant bits; the mantissa is a number indicating where the binary point is located and can be viewed as a scaling factor. Floating point arithmetic is typically used in engineering applications. In the early 1980s, it became possible to do floating-point arithmetic on a single chip by itself. Today, the floating-point arithmetic unit is only a small part of a general-purpose microprocessor.

    截至 2000 年 7 月,Adobe 的市值超过 160 亿美元,而 Xerox 的市值略高于 130 亿美元。

    As of July 2000, Adobe is worth over $16 billion, while Xerox is worth just over $13 billion.

    当时,作为一名贫困的研究生,我发现“免费”部分很重要。

    Being a poor graduate student at the time, I found the “free” part important.

    IC 中的驱动器使用相对较大的晶体管,通常有两种类型:双极型和 FET。双极晶体管通常速度更快,因为它们具有更高的电流驱动能力。要全面解释什么是双极晶体管和 FET 晶体管会占用太多篇幅。

    Drivers in IC use relatively large transistors, usually of two types: bipolar and FET. Bipolar transistors are generally faster due to their higher current driving capability. To fully explain what bipolar and FET transistors are would take up too much space.

    位是一个二进制数字,要么是 1,要么是 0。

    A bit is a binary digit, either 1 or 0.

    “Ply”是计算机国际象棋术语。这意味着两名玩家之一采取的行动。如果不是因为在国际象棋文献中,“走棋”可能意味着一对棋子即白棋和黑棋各走一步,那么“走棋”这个词可能更合适。

    “Ply” is computer chess jargon. It means a move made by one of the two players. The word “move” might have been more appropriate were it not for the fact that in chess literature a “move” could mean a pair of plies, with both White and Black making one move each.

硅芯片通常通过其封装的引脚    与外界进行通信。对于一些较新的 IC 封装,可能会使用焊盘。

    Silicon chips usually communicate with the outside world through the pins of their packages. With some of the newer IC packages, pads might be used instead.

    下面是优先级解码器的简单解释。六到六十四优先级解码器计算具有 6 位输入和 64 位输出的函数。6位输入N可以看作是从0到63的整数。如果64个输出位编号从0到63,则编号在N以下的所有输出位都应该是0,并且所有其他输出位都应该是1。

    Here is a simple explanation of a priority decoder. A six to sixty-four priority decoder computes a function with a 6-bit input and a 64-bit output. The 6-bit input N can be viewed as an integer from 0 to 63. If the 64 output bits are numbered from 0 to 63, then all the output bits numbered below N should be Os and all the other output bits should be 1s.

   “动态”对于芯片设计者来说具有特殊的意义。动态电路的状态或存储器以电荷的形式存储在电容器上,必须不断刷新。

   “Dynamic” has special meanings for chip designers. The states or memories of a dynamic circuit are stored as charges on capacitors, which have to be refreshed constantly.

   MOSIS(MOS [金属氧化物半导体]实施服务)是针对 VLSI 电路的低成本原型设计服务。最初,MOSIS 只为政府机构和大学提供服务。如今,MOSIS 还为商业公司提供服务。我在卡内基梅隆大学的时候,几乎所有美国大学设计的芯片都经过MOSIS。这些芯片实际上是免费的,因为 DARPA 和 NSF(国家科学基金会)等政府机构承担了成本。

   MOSIS (MOS [Metal-Oxide-Semiconductor] Implementation Service) is a low-cost prototyping service for VLSI circuits. Initially, MOSIS served only government agencies and universities. Today, MOSIS also provides services to commercial firms. While I was at Carnegie Mellon, almost all the chips designed at US universities went through MOSIS. The chips were effectively free as government sources such as DARPA and NSF (National Science Foundation) bore the cost.

   计算机国际象棋中的强力搜索意味着逻辑地检查每一个可能的走法,而选择性搜索则可能会删除一些走法。

   Brute force searching in computer chess means logically examine every possible move, in contrast to selective searching where some of the moves might be pruned away.

   选择性搜索对于不同的人来说意味着不同的事情。一般来说,选择性搜索有两种风格。一是更深层次地搜索“有趣”的线路(也称为选择性加深)。另一种是对“不感兴趣”的行进行浅搜索,或者根本不搜索,即剪掉“不感兴趣”的行(也称为选择性剪枝)。可以在同一个程序中混合两种口味。有些人认为选择性深化是修改后的蛮力搜索,而不是选择性搜索。

   Selective searching means different things to different people. In general, selective searching has two flavors. One is to search “interesting” lines deeper (also known as selective deepening). The other is to search “uninteresting” lines shallower, or not at all, that is, to prune away the “uninteresting” lines (also known as selective pruning). It is possible to mix the two flavors in the same program. Some people consider selective deepening as modified brute force searching and not selective searching.

   AB* 版本的 Hitech 出现于 20 世纪 90 年代初。它的公开锦标赛成绩并不比暴力版的 Hitech 好。

   A B* version of Hitech appeared in early 1990s. Its public tournament results were no better than those of the brute force version of Hitech.

 

 

第 4 章

CHAPTER 4

并非如此的国际象棋机器

The Chess Machine That Wasn’t


七周

Seven Weeks

托马斯·阿南塔拉曼 (Thomas Anantharaman) 在创建一个非常快的国际象棋程序方面的成就并没有在系里被忽视。1986 年最快的国际象棋程序是 Belle、Hitech 和 Cray Blitz。Belle 和 Hitech 是专用国际象棋机,大小与办公室冰箱差不多。Cray Blitz 是一个在价值数百万美元的 Cray 超级计算机上运行的国际象棋程序。当时的顶级程序每秒搜索 100,000 个国际象棋棋局。托马斯使用一个普通的办公工作站加上一个平装小说大小的国际象棋芯片接口卡,每秒可以得出 30,000 个国际象棋棋局,尽管国际象棋评估的大脑受到了损害

Thomas Anantharaman’s exploit in creating a very fast chess program did not go unnoticed in the Department. The fastest chess programs in 1986 were Belle, Hitech and Cray Blitz. Belle and Hitech were special purpose chess machines roughly the size of an office refrigerator. Cray Blitz was a chess program running on a multi-million-dollar Cray supercomputer. The top programs at the time were searching in the range of 100,000 chess positions per second. Thomas was getting 30,000 chess positions per second with a vanilla office workstation plus a chess chip interface card about the size of a paperback novel, although with a brain damaged chess evaluation.


频繁的 TG(感谢上帝,今天是星期五)派对是作为该系研究生的一些非正式的附带福利。它们并不总是在周五举行,并且可能出于各种原因举行。系里赞助了其中一些,但研究生和教员赞助了至少一半1

The frequent TG (Thank God [It’s Friday]) parties were some of the unofficial fringe benefits of being a graduate student in the Department. They were not always held on Friday, and could be held for all sorts of reasons. The Department sponsored some of them, but the graduate students and faculty members sponsored at least half1.

正是在其中一次 TG 上,与 Berliner 博士一起从事 Hitech 工作的研究生 Gordon Goetsch 开玩笑地建议我和 Thomas 参加即将在达拉斯举行的年度 ACM(计算机协会)计算机国际象棋锦标赛。戈登认为尽管有相反的说法,但锦标赛中的程序几乎没有国际象棋知识,而像托马斯这样的快速而愚蠢的程序将能够造成严重损害。它不会获胜——有更快的程序,但也没有那么愚蠢——但看看会发生什么会很有趣。当时托马斯和我都没有认真对待戈登的建议。

It was at one of these TGs that Gordon Goetsch, a graduate student working with Dr Berliner on Hitech, jokingly suggested that Thomas and I enter the program into the annual ACM (Association for Computing Machinery) Computer Chess Championship upcoming in Dallas. Gordon was of the opinion that the programs in the tournament, despite claims to the contrary, had very little chess knowledge, and a fast and dumb program, such as Thomas’, would be able to do serious damage. It would not win—there were faster programs that were not as dumb—but it would be fun to watch what would happen. Neither Thomas nor I took Gordon’s suggestion seriously at the time.

TG 结束后,Andreas Nowatzyk 和我与一些工程师一起来到了工程实验室。我并不打算按照戈登的建议去做,但我对国际象棋筹码有一个挥之不去的问题。托马斯的程序以每秒 30,000 个棋局的速度运行,我知道国际象棋芯片可以工作,但我不知道它可以以多快的速度运行而不会过热。如果我坚持最初的构建国际象棋芯片组的计划,那么至少还要过一年才能知道答案。这将是一个冒险的提议。我可能应该构建一些可以以更高的速度运行国际象棋筹码的东西。如果我建造了这样一个速度测试仪,那么将其修改为“穷人的棋机”并参加ACM比赛将是相当有趣的。当然,也存在一些障碍。

After the TG, Andreas Nowatzyk and I ended up in the Engineering Lab with some of the staff engineers. I did not intend to do what Gordon suggested, but I had a lingering question about the chess chip. With Thomas’ program running at 30,000 chess positions/sec, I knew that the chess chip worked, but I had no idea at what speed it could run without overheating. If I stuck with my original plan of building a chess chip set, it would be at least another year before I knew the answer. That would be a risky proposition. I probably should build something that could exercise the chess chip at a higher speed. If I built such a speed tester, then modifying it into a “poor man’s chess machine” and entering the ACM competition would be quite interesting. Of course, there were a few obstacles.

首先,距离ACM比赛只剩下七周了。其次,我只有学生预算,即什么都没有。最后但并非最不重要的一点是,如果我想参加七周后的 ACM 锦标赛(有竞争性项目),我需要帮助。

First, there were only seven weeks left before the ACM competition. Second, I only had a student budget, namely, nothing. Last but not least, I needed help if I wanted to participate in the ACM Tournament, with a competitive program, in seven weeks time.

当我向安德烈亚斯和工程师提到速度测试仪的想法时,他提供了第一个帮助。他的眼睛一亮,并提出了一个获取原型板的来源,该原型板可用于制造测试仪——Perq 工作站的部门废料堆。第一个进入市场的计算机工作站 Perq 既超前又超越了时代。工作站市场最终起飞,但 Perq 仍坚持旧的多芯片 CPU(中央处理器单元)设计,被使用更便宜的单芯片微处理器 CPU 的更灵活的竞争对手所取代。Perq 的制造商三河计算机公司 (Three River Computers) 于 1986 年停业,该部门淘汰了大部分 Perq。废料堆里充满了 Perq 原型板,为其中一块配备绕线插座作为全尺寸绕线板的替代品将是一件简单的事情。借给我总线接口卡的工程师劳伦斯·布彻 (Lawrence Butcher) 自愿在设计完成后对电路板进行绕线。还有一个问题——获取电子设备测试仪的零件。孔教授的柜子里有很多用于 DARPA 项目的备件。我决定从柜子里借一些零件——如果需要的话,可以随时从测试板上回收。那天晚上,我拿了 Kung 柜子里的备件清单,Andreas 找到了一块形状不错的 Perq 原型板。测试仪将只使用孔教授柜子里的零件来设计,尽管他直到芯片测试仪完成后才知道借来的零件。

Andreas provided the first piece of help when I mentioned the idea of the speed tester to him and the staff engineers. His eyes lit up and he suggested a source for getting a prototype board that could be used to make the tester—the departmental scrap heap for Perq workstations. The first computer workstation to hit the market, the Perq, was both before its time and beyond its time. The workstation market eventually took off, but the Perq, stuck with an older multi-chip CPU (central processor unit) design, was taken over by nimbler competitors that used cheaper single-chip microprocessor CPUs. Three Rivers Computers, the manufacturer of Perq, went out of business in 1986, and the Department retired most of the Perqs. The scrap heap was full of Perq prototype boards, and it would be a simple matter to outfit one of them with wire-wrap sockets as a substitution for a full-sized, wire-wrap board. Lawrence Butcher, the engineer who lent me the bus interface card, volunteered to wire wrap the board when the design was complete. There was one more problem—getting the electronic parts for the tester. Professor Kung had a lot of spare parts in his cabinet for a DARPA project. I decided to borrow some parts from the cabinet—if they ever became needed, they could always be reclaimed from the tester board. That night, I took an inventory list of the spare parts in Kung’s cabinet, and Andreas located a Perq prototype board in decent shape. The tester would be designed using only the parts in Professor Kung’s cabinet, although he did not know about the borrowed parts until after the chip tester was completed.

第二天早上,我和托马斯聊了聊。我向他介绍了芯片测试仪的情况,新的软件界面会是什么样子,会提供什么样的新硬件功能。最重要的新功能是能够直接在测试器硬件中进行搜索,这将使我们能够更快地搜索并以更高的速度运行国际象棋芯片。考虑到设计时间的限制,在硬件中执行的搜索必然是简单的。硬件搜索不会知道易位走法或升级不足走法(将棋子升级为车、象或马,而不是像正常典当升级中的后)。也就是说,硬件甚至不知道所有合法的国际象棋走法。此外,硬件不会知道重复一个位置可能会导致国际象棋中的平局。此外,由于它只是一个筹码测试器,因此硬件将仅根据棋子放置表(列出棋盘每个方格上每个棋子的值的表格)来评估国际象棋位置。它可以用作国际象棋机器,但非常粗糙。为了弥补硬件限制,搜索过程中沿任意线检查的前几层将在软件中完成,从而了解所有合法的移动和重复抽牌。最后几层将在硬件中完成。总体速度将接近硬件速度,因为距根2位置 10 层的新位置比距根 2 位置 6 层的新位置多得多。还可以在搜索树的软件部分添加更复杂的国际象棋评估。

The next morning, I had a chat with Thomas. I briefed him on the chip tester, what the new software interface would look like, and what kind of new hardware functions would be provided. The most important new function was the ability to do searches directly in the tester hardware, which would then allow us to search faster and exercise the chess chip at higher speed. The searches performed in hardware would necessarily be simple minded, given the design time constraint. The hardware searches would not know about either castling moves or underpromotion moves (promoting a pawn to a rook, a bishop or a knight, instead of a queen as in a normal pawn promotion). That is, the hardware would not even know all the legal chess moves. Moreover, the hardware would not know that repeating a position could lead to a rule draw in chess. Also, since it was going to be just a chip tester, the hardware would evaluate a chess position based only on a piece-placement table (a table tabulating the value of every piece on each square of the chessboard). It could be used as a chess machine, but a very crude one. To compensate for the hardware limitation, the first few plies examined along any line during the search would be done in software, thus knowing about all the legal moves and the repetition draws. The last few plies would be done in hardware. The overall speed would be close to the hardware speed, since there are far more new positions, say, 10 plies away from the root2 position than new positions, say, six plies away. It was also possible to add more complicated chess evaluation in the software part of the search tree.

我们需要对托马斯的计划做出改变比较简单,他就同意做。托马斯是一位优秀的程序员,但我们都不是国际象棋棋手。我们至少需要一名像样的国际象棋棋手来编写评估软件并准备开局书(计算机在游戏开始时会下的棋步)。

The changes that we needed to make to Thomas’ program were relatively simple, and he agreed to make them. Thomas was a good programmer but neither of us was a chess player. We needed at least one decent chess player to write the evaluation software and to prepare the opening book (the moves that the computer would play at the beginning of the game).

我试图让迈克·布朗再次感兴趣。他是我认识的最好的程序员之一,他在高中时曾参加过国际象棋比赛,但他专注于其他事情。系里还有几位研究生,棋艺很好。其中一位彼得·詹森(Peter Jansen)表达了一些兴趣,但他正忙于课程作业。

I tried to get Mike Browne interested again. He was one of the best programmers that I knew, and he had played chess competitively in high school but he was preoccupied with other things. There were a few other graduate students who were good chess players in the Department. One of them, Peter Jansen, expressed some interest, but he was busy with course work.

默里·坎贝尔(Murray Campbell)是托马斯硬件辅助程序的第一个受害者,他可能是研究生中最强的国际象棋棋手,而且他以前曾研究过计算机国际象棋。穆雷是一位国际象棋神童,曾两次代表阿尔伯塔省参加加拿大青少年锦标赛。当他还在参加最后一次国际象棋锦标赛时,当他还在寻找奖金时,他同意提前抽签,这样他就可以打网球,因此他放弃了认真的国际象棋。用他的话说,“从那时起我就知道我不想再下棋了。” 他本来可以成为球队的重要补充,但有一个复杂的问题。他是海泰集团的,我去追他不太合适。

Murray Campbell, the first victim of Thomas’ hardware-assisted program, was perhaps the strongest chess player among the graduate students, and he had worked on computer chess before. Murray was a chess prodigy, twice representing the Province of Alberta in the Canadian Junior Championship. He had quit playing serious chess when, in his last chess tournament while he was still in the prize hunt, he agreed to an early draw so that he could play tennis. In his words, “That was when I knew that I didn’t want to play competitive chess any more.” He would have been a great addition to the team, but there was one complication. He was in the Hitech group, and it would not be proper for me to go after him.

在工程实验室的谈话两周后,我意外地收到了默里发来的一封电子邮件,询问测试板上的棋子放置台是否可以在每次移动后进行调整。如果是的话,假设我们感兴趣,他对如何在前几次移动之后使程序进入合理的位置有一些想法。桌子确实是可以调整的,在发现 Hitech 可能不会参加锦标赛后,我们接受了他的提议。此时,我向 ACM 锦标赛委员会提交了参赛作品,但尚未完成设计,也没有工作计划。我们将新程序称为 ChipTest,部分是因为硬件的芯片测试器性质,部分是因为我们想强调它不是真正的国际象棋机器。

Two weeks after the conversation in the Engineering Lab, I got a surprise e-mail from Murray asking whether the piece-placement table on the tester board was adjustable after each move. If it was, he had some ideas on how to get the program into reasonable positions after the first few moves, assuming that we were interested. The table was indeed adjustable and, after finding out that Hitech was probably not going to play in the tournament, we took him up on his offer. I submitted the entry, without a finished design and without a working program, to the ACM tournament committee at about this point. We called the new program ChipTest, in part because of the chip tester nature of the hardware, and in part because we wanted to stress that it was not a real chess machine.

七周后的第四周,硬件设计开始成形。我感觉很舒服,可以公开露面,并发布了一条电子消息寻求更多帮助。由于穆雷没有承诺研究评估功能,我与其他几位国际象棋棋手进行了交谈,但没有成功。然而,我确实得到了盖伊的积极回应雅各布森。盖伊不仅仅是一名普通的研究生。他是 Rog-O-Matic 的作者之一,这是一款玩 Rogue 的自动程序,Rogue 是当时研究生中流行的电脑游戏。众所周知,他是该部门最优秀的程序员之一。

Four weeks into the seven-week period, the hardware design began taking shape. I felt comfortable enough to go out in the open and posted an electronic message seeking additional help. Since Murray had not promised to work on the evaluation function, I talked with a few other chess players but with no luck. However, I did get one positive response to my message from Guy Jacobson. Guy was not just a garden-variety graduate student. He was one of the authors of Rog-O-Matic, an automatic program that played rogue, a popular computer game among graduate students at the time. By all accounts, he was one of the premier programmers in the Department.

托马斯、穆雷、盖伊和我开了一个会。托马斯将完成大部分编程工作。穆雷将承担准备开局书和编写棋子排列表软件的任务,这实际上是主要的评估职能。当确定 Hitech 不会参加即将到来的比赛时,穆雷扩大了他的参与范围。盖伊将致力于对棋子结构进行软件评估,以便程序知道棋子在搜索树的软件部分中有多强或多弱。

Thomas, Murray, Guy, and I had a meeting. Thomas would do the bulk of the programming. Murray would take up the tasks of preparing the opening book and writing software for the piece-placement table, which was effectively the main evaluation function. Murray expanded his involvement when it became certain that Hitech would not play in the upcoming competition. Guy would work on a software evaluation for pawn structure so that the program would know how strong or weak the pawns were in the software part of the search tree.

第五周结束时,硬件已准备好进行绕线,劳伦斯·布彻 (Lawrence Butcher) 一夜之间完成了这项原本需要我几天时间的工作。硬件大约有 90 个芯片,如果我们必须自己购买这些部件,成本约为 500 美元。

The hardware was ready for wire wrapping by the end of the fifth week, and Lawrence Butcher completed overnight the job which would have taken me a few days. The hardware had about ninety chips which would have cost about $500 if we had had to acquire the parts ourselves.

由于时间有限,我以极简的态度设计了测试仪。不幸的是,我走得太远了——测试仪完全不包含用于测试测试仪本身的电路。然而,测试仪是可微编程的;也就是说,可以用微程序对其进行编程,直接控制所有硬件电路。理论上,假设我有测试微程序,测试器可以在某种程度上进行测试。但为了验证测试微程序的正确性,我需要有工作的硬件。这是经典的先有鸡还是先有蛋的问题。

Because of the time constraint, I designed the tester with a minimalist attitude. Unfortunately, I went too far—the tester contained absolutely no circuitry for testing the tester itself. The tester was, however, microprogrammable; that is, it could be programmed with a microprogram, which controls all the hardware circuits directly. In theory, the tester could be tested to some extent, assuming that I had the testing microprograms. But in order to verify the correctness of the testing microprograms, I needed to have working hardware. It was the classic chicken and egg problem.

安德烈亚斯通过创建一个高质量的微汇编器(一种将微程序转换为适当的位模式来控制硬件的程序)来拯救我,这使得微程序的编码变得更加容易。仍然花了大约一周的时间和几个微程序来完成硬件的测试。还需要编写一个微程序——控制硬件搜索功能的微程序。时间越来越短了。

Andreas came to my rescue by creating a high quality microassembler (a program that converts the microprograms into appropriate bit patterns for controlling the hardware) that made the coding of the microprograms much easier. It still took about one week and several microprograms to complete the testing of the hardware. There was one more microprogram to write—the one that controlled the hardware search function. Time was getting short.

与此同时,盖伊没有抽出时间来开发该程序,这意味着我们不会对典当结构进行软件评估。哦,好吧,盖伊的帮助本来就是一份礼物。

Meanwhile, Guy did not get around to work on the program which meant we would not have a software evaluation for pawn structure. Oh, well, Guy’s help was a gift to begin with.

现在还不是恐慌的时候,但我们需要制定应急计划。没有硬件搜索我们仍然可以玩,但是我们将遭受严重的速度损失。搜索微程序比所有测试微程序的总和要大得多,而且调试起来也困难得多。出发前往 ACM 锦标赛举办地达拉斯的前一天晚上,我没有睡觉,而是继续工作,直到我必须登机。安德烈亚斯开车送我去机场,一路上我都保持沉默。比赛进行期间,托马斯会接手微程序的工作。我们没有一个适用于整个比赛的微程序。

It wasn’t time to panic yet, but we needed to have a contingency plan. We could still play without the hardware search, but we would suffer a heavy speed penalty. The search microprogram was much bigger than all the test microprograms combined, and it was also far more difficult to debug. I did not sleep the night before I was to leave for Dallas, site of the ACM tournament, but continued working until I had to get on the plane. Andreas drove me to the airport, and I was silent the entire way. Thomas would pick up the work on the microprogram while the competition was going on. We did not have a working microprogram for the entire tournament.

去年,我邀请肯·汤普森加入我的论文委员会,但我只是在达拉斯才第一次见到他。他作为赛事官员参加了比赛,但在克雷·布利茨(Cray Blitz)以机器时间不足为由退出后,贝尔被选中参加比赛,而肯则成为了参赛者之一。那一年的头号种子是贝贝,我们第一轮就和贝贝队配对。《贝贝》的作者是来自芝加哥地区的一对夫妇托尼·谢尔泽 (Tony Scherzer) 和琳达·谢尔泽 (Linda Scherzer)。我和托尼的第一次谈话进行得不太顺利。我试图解释为什么 ChipTest 不会全速,告诉托尼我们在比赛前七周就开始构建它,并且我们还需要一两周才能达到最高速度。托尼说:“是的,他们都说要休息一周。” 对他来说,我听起来肯定像个吹牛者(在电脑国际象棋世界里这样的人有很多)。Bebe 是另一台专用国际象棋机器,事实证明它是更强大的玩家,ChipTest 没有经过太多战斗就失败了。当我更加了解托尼时,我发现他是一个好人。他对自己所做的事情非常开放,我与他进行了许多互惠互利的对话。

The previous year I had asked Ken Thompson to be on my thesis committee, but I only met him for the first time in Dallas. He was there as a tournament official, but after Cray Blitz pulled out, citing unavailability of machine time, Belle was drafted to play and Ken became one of the competitors. The top seed that year was Bebe, and we were paired against the Bebe team in the first round. The authors of Bebe were Tony and Linda Scherzer, a couple from the Chicago area. My first conversation with Tony did not go too well. I tried to explain why ChipTest would not be at full speed by telling Tony that we had started to build it seven weeks before the tournament, and that we needed another week or two before we could get to the top speed. Tony said, “Yeah, they all said that they were off by a week.” To him, I must have sounded like a braggart (and there were plenty of them around in the computer chess world). Bebe, which was another special purpose chess machine, proved to be the stronger player and ChipTest went down without much of a fight. Tony turned out to be a nice guy when I got to know him better. He was quite open about what he was doing and I had many mutually beneficial conversations with him.

下一轮我们对阵 Monty Newborn 的鸵鸟队。Monty 从 1970 年开始组织 ACM 锦标赛,这是 Ostrich 的最后一次锦标赛。Ostrich 是早期的一个竞争者,但它不再被认为是一个强大的程序。ChipTest 在对阵 Ostrich 的比赛中相当轻松地取得了胜利,然后灾难就降临了。ChipTest 的软件和硬件仍然存在一些错误,Thomas 进行了“修复”。不幸的是,头脑简单的“治疗”结果比疾病更糟糕。每当程序检测到有问题时,它就会只执行硬件移动生成器生成的第一个移动,这实际上是随机选择的移动。在这场比赛中,ChipTest 打出了由于“治愈”的结果而出乎意料。这步棋牺牲了一个车来换取一个兵,而有了车,游戏就结束了。托马斯通宵工作,用更安全的替代方案取代了“治疗”,即记录以前的最佳走法,每当程序失控时,就执行记录的最佳走法。

In the next round we played Monty Newborn’s Ostrich. Monty organized the ACM tournaments from their beginnings in 1970, and this was Ostrich’s last tournament. Ostrich was a contender in earlier days, but it was no longer considered a strong program. ChipTest got a winning position fairly easily against Ostrich, and then disaster struck. There were still a few bugs in both ChipTest’s software and hardware, and Thomas put in a “cure”. Unfortunately, the “cure”, being very simple minded, turned out to be worse than the disease. Whenever the program detected that it was having problems, it would just play the first move generated by the hardware move generator, which was effectively a randomly chosen move. In this game, ChipTest played a move out of the blue as the result of the “cure”. The move sacrificed a rook for a pawn, and with the rook, the game. Thomas worked overnight and replaced the “cure” with a safer alternative, that is, to record what was previously the best move and whenever the program ran amok, play the recorded best move.

ChipTest在接下来的比赛中毫不费力地战胜了Rex,取得了首场胜利。下一轮,它对阵的是来自德国的Merlin,一个备受推崇的节目。ChipTest 通过搜索超越 Merlin 赢得了比赛,但也发生了一些事件。它打出了非法的易位棋,但同时也表示该易位棋是非法的,并且程序处于不一致的状态。赛事总监和我们的对手同意让我们重新开始比赛。重新开始后,它下了合法的一步,ChipTest继续赢得了比赛。

ChipTest won the next game against Rex without much difficulty, and recorded its first win. The next round, it played Merlin from Germany, a well-respected program. ChipTest won the game by outsearching Merlin, but not without some incidents. It played an illegal castling move, but it also stated that the castling move was illegal and that the program was in an inconsistent state. The tournament director and our opponent agreed to let us restart the program. It played a legal move after the restart, and ChipTest went on to win the game.

上一轮,ChipTest对阵Recom。Recom 是 Rebel 的前身,Rebel 是当今最好的商业国际象棋程序之一,即使在 1986 年,它也是一个强大的程序。ChipTest 早期运行良好,但 Ostrich 游戏的问题再次出现。有了新的“治愈”,它的招式并没有被直接杀死,但位置却变得非常糟糕。

In the last round, ChipTest played against Recom. Recom was a predecessor of Rebel, one of the best commercial chess programs today, and it was a strong program even back in 1986. ChipTest was playing well early on, but the problems from the Ostrich game resurfaced. With the new “cure”, it played moves that did not get killed outright, but the position was getting really bad.

然后奇迹发生了。ChipTest突然声称比赛已经平局,并且下了棋,牺牲了另一辆车换了一个棋子。但与鸵鸟游戏中的牺牲不同,这是一个很好的牺牲。我兴奋得上蹿下跳。过了一会儿,雷科姆意识到比赛确实是平局,我们宣布休战。此后每次见到Recom的操作员,他总会提起游戏最后我上蹿下跳的场景。

Then a miracle happened. ChipTest suddenly claimed that the game was drawn, and it played a move that sacrificed another rook for a pawn. But unlike the sacrifice in the Ostrich game, this was a good sacrifice. I got so excited that I was jumping up and down. Recom, after a while, realized that the game was indeed drawn, and we called the truce. Every time that I met Recom’s operator since, he always brought up this scene of my jumping up and down at the end of this game.

由于某种未知的原因,Recom 游戏的官方公告中发布了错误的游戏分数。Berliner 博士发现发布的最终位置对 ChipTest 来说是一场彻底的损失,他很惊讶地给我发电子邮件询问 Recom 是否崩溃了。似乎没有人发现这个错误。真实的比赛得分在附录 B中给出,由 Ken Thompson 提供。

For some unknown reason, the wrong game score was posted for the Recom game in the official bulletin. Dr Berliner found the posted final position a dead loss for ChipTest and was surprised enough to e-mail me asking whether Recom had crashed. No one else seemed to have caught the mistake. The real game score is given in Appendix B, courtesy of Ken Thompson.

ChipTest的首次亮相算不上是一场出柜派对,但第一次尝试的平局(两胜两负一平)也算不错的结果。

ChipTest’s debut was not quite a coming-out party, but an even score (two wins, two losses and one draw) on the first try was not exactly a bad result.

单一扩展

Singular Extensions

尽管贝贝是1986年ACM锦标赛的头号种子,但他并没有赢得冠军。鉴于使用“瑞士”锦标赛形式,这并不令人意外。在瑞士的锦标赛中,分数相似但尚未交手的选手会被配对参加下一轮比赛。瑞士锦标赛比​​淘汰赛更好,因为有一些机会从早期的失利中恢复过来,但顶级选手获胜的几率仍然不大。最强的程序赢得 ACM 或类似锦标赛的可能性实际上非常小。

Despite being the top seed at the 1986 ACM tournament, Bebe did not win the Championship. Given the use of the “Swiss” tournament format, this was not really a surprise. In a Swiss tournament, players with similar scores who have not yet played against each other are paired together for the next round. A Swiss tournament is better than a knockout tournament in that there is some chance to recover from an early loss, but the odds are still not good for the top player to win. The probability of the strongest program winning the ACM or a similar tournament is actually quite slim.

1986年,贝贝输给了最终的冠军贝拉。在第二名的争夺中,Bebe 被 Lachex(洛斯阿拉莫斯国际象棋实验)击败,Lachex 是一个在洛斯阿拉莫斯国家实验室的 Cray 超级计算机上运行的程序。托尼·谢尔泽对这次失利感到非常失望。他告诉每个愿意听的人游戏中发生的事情,这是一个有趣的故事。

In 1986, Bebe lost to Belle, the eventual winner. In the battle for second place, Bebe was defeated by Lachex (Los Alamos CHess EXperiment), a program running on a Cray supercomputer at the Los Alamos national lab. Tony Scherzer was quite disappointed about the loss. He told everyone who would listen what had happened in the game, and it was an intriguing story.

贝贝和拉切克斯不知何故陷入了一系列被迫的行动,双方一路上只有一个好的选择。这两个项目都不知道谁最终会取得领先。双方都继续玩“唯一”的棋子,直到两个程序突然意识到贝贝彻底输了。Bebe 是顶级程序之一,Lachex 也毫不逊色。这两个程序怎么会这么无能呢?对于这个问题我们可以做些什么呢?

Bebe and Lachex somehow went into a sequence of forced moves where both sides had only one single good choice along the way. Neither program had any idea about who would come up ahead in the end. Both sides just kept on playing the “only” moves until, suddenly, both programs realized that Bebe was a dead loss. Bebe was one of the top programs, and Lachex was no slouch either. How could both programs be so clueless? And what could be done about the problem?

ACM 锦标赛不仅仅是一场比赛;更是一场比赛。这也是计算机国际象棋研究人员的年度聚会。在回合之间,每当一小群人聚在一起时,就会进行即兴讨论。我与 Tony Scherzer 和其他一些人进行的讨论之一是关于选择性搜索的想法。托尼笼统地表示,“选择性剪枝”(将“没有希望”的棋步从考虑中剔除)的旧理念已经死了,取而代之的是“选择性深化” 3或“选择性延伸”的新理念,即更深入地寻找有趣的棋

The ACM tournament was not just a competition; it was also the annual gathering of computer chess researchers. Between rounds, impromptu discussions took place whenever a small group of people got together. One of the discussions that I had with Tony Scherzer and a few others was about the idea of selective search. Tony made a blanket statement that the old idea of “selective pruning” (discarding “unpromising” moves from consideration) was dead, replaced by the new idea of “selective deepening”3 or “selective extensions”, namely, searching interesting moves more deeply.

“选择性深化”的想法可以用来解决Lachex-Bebe博弈中出现的问题吗?好吧,如果我们有某种方法来检测当前位置只有一个好棋,那就可以了。我想了想,得出的结论是,没有一种确定的方法可以不付费就获得这些信息。在一些非常罕见的情况下,信息是免费的,但我们谈论的可能是百万分之一。我放这个问题一直萦绕在我的脑海里,直到第二天我登上返回匹兹堡的飞机。

Could the idea of “selective deepening” be used to solve the problem that showed up in the Lachex-Bebe game? Well, it could if we had some way to detect that there was only one good move for the present position. I thought about it briefly, and reached the conclusion that there was no sure way to obtain the information without paying for it. In some very rare situations, the information is free, but we were talking about perhaps once in a million. I put the question in the back of my mind until the next day when I got on the plane back to Pittsburgh.

人们进行选择性深化或选择性扩展的方式大多是基于国际象棋知识。如果一侧受到限制,请更深入地检查该位置。如果一个棋子即将达到第八级并成为皇后,则更深入地搜索该位置。就 ChipTest 而言,很难将国际象棋知识融入到更具选择性的扩展中。但是,如果我们能让微程序运行,我们将拥有地球上最快的国际象棋程序,而且搜索也会相对便宜。至少“单一扩展”的想法就是这样开始的。确实没有必要有一个快速的程序。无论如何,我认为我们可以通过修改搜索或进行额外的搜索来收集有关移动是否单一或唯一合理的移动的信息,从而稍微减慢程序速度。一旦我们验证了这一举动是奇异的,那么我们就可以更深入地搜索它。收集奇点信息的一种方法是显式搜索所有替代移动并测试它们是否比所讨论的移动明显更差。这可以沿着“主要变化”或 PV(双方发挥最佳动作的线)相当有效地完成,因为无论如何都必须在正常搜索中检查沿着 PV 的所有替代动作。这一基本思想后来成为“PV奇异延伸”。不管怎样,我当时认为这个想法只是理论上的好奇心。回到匹兹堡后,我还有更重要的事情要处理。是时候提交我的论文提案了。

The way people had been doing selective deepening or selective extensions was mostly based on chess knowledge. If one side is in check, examine the position deeper. If a pawn is about to reach the eighth rank and become a queen, search the position deeper. In the case of ChipTest, it was very difficult to put in the chess knowledge for more selective extensions. But, if we could get the microprogram working, we would have the fastest chess program on the planet, and searching would be comparatively cheap. At least that was how the idea of “singular extensions” began. It was not really necessary to have a fast program. In any case, I reasoned that we could afford to slow down the program a little bit by modifying the search or doing extra searches to gather information about whether a move is singular, or the only reasonable move. Once we verified that the move is singular, then we could search it more deeply. One way to gather the singularity information is to explicitly search all the alternative moves and to test whether they are significantly worse than the move in question. This could be done fairly efficiently along the “principal variation”, or PV (the line where both sides play the best moves), as all the alternative moves along the PV have to be examined in a normal search anyway. This basic idea later became the “PV singular extensions”. Anyway, I deemed the idea a theoretical curiosity at the time. I had more important things to attend to when I got back to Pittsburgh. It was about time to present my thesis proposal.

尽管肯·汤普森(Ken Thompson)在我的论文委员会工作了一年多,但我还没有提出我的论文提案。当我完成国际象棋走棋生成器芯片的设计后,Kung教授建议Berliner博士加入我的委员会。当时这似乎是个好主意,如果我想最终建造出终极的国际象棋机器,我需要在未来的某个时候至少有一位优秀的国际象棋棋手的合作。柏林纳博士无疑是卡内基梅隆大学最好的国际象棋棋手,也是世界函授国际象棋(邮寄国际象棋)冠军。他的世界冠军并不意味着他在常规国际象棋比赛中处于顶尖梯队,但他仍然具有国际大师4级的实力,这将使他跻身美国前五十名棋手之列。成为如此强大的玩家本身就是一项相当大的成就。他可以同时和二十个和我实力相当的选手对战,并且能赢下每一场比赛。讽刺地,Berliner 博士加入委员会可能就注定了任何合作的可能性5。不管怎样,我采纳了孔教授的建议,Berliner博士成为了我的论文委员会的第三位成员。VLSI 小组的教授 Alan Fisher 是第四位成员。

Even though Ken Thompson had been on my thesis committee for over a year, I had not yet presented my thesis proposal. After I finished the design of the chess move generator chip, Professor Kung suggested that Dr Berliner should be on my committee. It seemed a good idea at the time that if I wanted to eventually build the ultimate chess machine, I would need the cooperation of at least one good chess player at some future date. Dr Berliner was without doubt the best chess player at Carnegie Mellon as well as a World Correspondence Chess (chess played by mail) Champion. His world title did not mean that he was of the top echelon in regular chess, but he was still of International Master4 strength, which would have placed him among the top fifty over-the-board players in USA. To be such a strong player is a considerable achievement in itself. He could play simultaneously against twenty players of my strength and win every single game. Ironically, having Dr Berliner on the committee probably doomed any possibility of cooperation5. Anyway, I took Professor Kung’s suggestion, and Dr Berliner became the third member of my thesis committee. Alan Fisher, a professor in the VLSI group, was the fourth member.

当我从 ACM 锦标赛回来时,托马斯已经清理并重写了搜索微程序。它可以工作,但尚未优化,并且仍然存在一些硬件错误,不时导致微程序崩溃。在准备论文提案时,我断断续续地研究硬件和微程序。ACM 锦标赛一个月后,我提出了我的建议,即完成一台真正的国际象棋机器的设计,并制定出如何使用多个国际象棋机器以达到前所未有的搜索速度的细节。这篇论文完成后,成为了未来深蓝国际象棋机器的蓝图。

When I got back from the ACM Tournament, Thomas had cleaned up and rewritten the search microprogram. It was working but not yet optimized, and there were still some hardware bugs that crashed the microprogram from time to time. I worked on both the hardware and the microprogram off and on while preparing for my thesis proposal. A month after the ACM tournament, I presented my proposal, which was to complete the design of a real chess machine and work out the details of how to use multiple chess machines to get to unprecedented search speed. The thesis, when it was finished, became the blueprint of the future Deep Blue chess machine.

提出建议后不久,我调试了ChipTest硬件并优化了微程序。托马斯也有一些空闲时间来开发该软件。ACM 锦标赛两个月后,ChipTest 与 Hitech 进行了第一场比赛。两队当时关系融洽,都有意找一位陪练搭档6。ChipTest 赢得了比赛,但总体而言 Hitech 可能仍然是更强的程序。然而,在首次亮相仅两个月后,ChipTest 显然已经成为世界顶级国际象棋机器之一。它每秒搜索大约 300,000 个国际象棋位置,大约比任何其他国际象棋程序快三倍。对于设计为芯片测试仪的东西来说还不错。

Shortly after the proposal, I debugged the ChipTest hardware and optimized the microprogram. Thomas also had some free time to work on the software. Two months after the ACM Tournament, ChipTest played its first game against Hitech. The two teams were on cordial terms at the time and both were interested in getting a sparring partner6. ChipTest won the game, but Hitech was probably still the stronger program overall. ChipTest, however, was clearly already one of the top chess machines in the world, just two months after its lackluster debut. It was searching about 300,000 chess positions per second, approximately three times faster than any other chess program. Not bad for something that was designed as a chip tester.

ChipTest成为稳定的棋机后,我进行了一些讨论与 Murray Campbell 讨论 PV 奇异扩展的想法。我自己并不是一个真正的国际象棋棋手,我以穆雷为参考,来看看这个想法是否有意义。默里认为它值得认真测试。托马斯在周末实现了这个想法,我们将带有 PV 奇异扩展的版本与没有它的版本进行了二十场比赛。结果是一个很大的惊喜。新版本以15比5战胜了旧版本。三比一的胜负比相当于200个等级分的差距,或者一个完整的国际象棋等级。这是关于世界冠军和典型特级大师之间的区别。它是巨大的。我们都很兴奋,尽管我们的兴奋因为知道比赛结果是在两个密切相关的节目之间进行而缓和,而且实力差异可能被大大夸大了。这二十场比赛是从十个不同的起始位置进行的,颜色交替。对于处于相同起始位置的一对游戏,通常两个程序会采取完全相同的动作,直到一个关键位置,新版本设法避免立即发生灾难并最终平局或赢得比赛。以《星际迷航》来类比,该计划的两个版本可以被视为两艘相同的企业级星舰,一艘所有远程传感器均被禁用,另一艘除了前面的远程传感器之外的所有传感器均被禁用。当两艘星际飞船都驶向一个巨大的黑洞时,装有工作前远程传感器的星际飞船将探测到黑洞,并及时逃脱其致命的控制。

After ChipTest became a stable chess machine, I had some discussions with Murray Campbell about the idea of PV singular extensions. Not being a real chess player myself, I used Murray as a sounding board to find out whether the idea made sense. Murray was of the opinion that it deserved serious testing. Thomas implemented the idea during a weekend, and we pitted the version with PV singular extensions against a version without it in a twenty-game match. The result was a big surprise. The new version beat the old version by 15 to 5. A win-loss ratio of three to one corresponds to a difference of 200 rating points, or a full chess class. It is about the difference between the World Champion and a typical Grandmaster. It is huge. We were all very excited, although our excitement was tempered by the knowledge that the match result was between two closely related programs, and the strength difference was probably greatly exaggerated. The twenty games were played from ten different starting positions with alternate colors. For the pair of games from the same starting position, usually the two programs would play exactly the same moves until a critical position where the new version managed to avoid an immediate disaster and ended up drawing or winning the game. Taking an analogy from ‘Star Trek’, the two versions of the program could be viewed as two identical Enterprise-class starships, one with all the long-range sensors disabled and one with all but the front long-range sensors disabled. As both starships head towards a massive black hole, the starship with the working front long-range sensor will be the one that detects the black hole, and escapes its deadly clutch in time.

任何有自尊心的星舰船长都不会允许他的飞船长期没有全套远程传感器。如果远处右舷有一颗有趣的行星怎么办?如果没有右舷远程传感器,我们将完全错过这颗行星。我们发现自己也处于类似的境地。PV奇异扩展使程序能够发现当前最佳计划的潜在灾难,但扩展并没有帮助程序找到当前计划的有趣替代方案。用国际象棋术语来说,PV 奇异扩展并不能帮助程序发现组合,即导致显着优势的非明显的移动序列。我们需要一个新想法。我们能否扩展 PV 奇异扩展的思想来解决这个问题?

No self-respecting starship captain would allow his ship to be without the full complement of long-range sensors for long. What if there was an interesting planet starboard in the distance? Without the starboard long-range sensor, the planet would be missed completely. We found ourselves in a similar situation. PV singular extensions allowed the program to discover the potential disaster with the current best plan, but the extensions did not help the program find interesting alternatives to the current plan. In chess jargon, PV singular extensions does not help the program discover combinations, that is, non-obvious sequences of moves that lead to significant advantages. We needed a new idea. Could we extend the idea of PV singular extensions to attack the problem?

如果替代的国际象棋走法不如我们当前的最佳走法,那么我们的对手必须对每一种可能的方法进行反驳我们可以继续玩。为了验证某步棋对我们来说不是最好的棋步,我们需要为对手找到一个有效的反驳棋步来对抗我们的棋步。如果我们的对手在我们的棋步之后只有一个有效的反驳棋步怎么办?那么,在对手独特的反驳举动之后进行更深入的搜索似乎是有意义的。直观地说,如果玩家只有一招好棋,那么该棋子很可能是唯一能推迟玩家不可避免的灾难的棋子这是我们所谓的高失败单一扩展背后的主要原因。(计算机国际象棋研究人员将反驳棋步称为高失败棋步。)

If an alternative chess move is not as good as our current best move, then our opponent must have refutations to every possible continuation that we could play. To verify that a move was not the best move for us, we needed to find one working refutation move for our opponent against our move. What if our opponent had only one working refutation move after our move? Then it seemed to make sense to search deeper after our opponent’s singular refutation move. Intuitively, if a player has only one good move, it is possible that the move is the only one left that delays the inevitable disaster for the player. This is the main reasoning behind what we called fail-high singular extensions. (Computer chess researchers refer to refutation moves as fail-high moves.)

高故障单一扩展的引入使 ChipTest 检测组合的能力实现了巨大飞跃。但在面对面的竞争中,最新版本并没有比旧版本表现得更好。高失败奇异扩展的额外计算成本消除了增益。我们决定保留新的扩展,因为我们认为它们可能会对人类玩家产生积极的影响。至少,这些扩展将使人类玩家无法猜测该程序是否会出现特别深度的组合。

The introduction of fail-high singular extensions produced a quantum leap in ChipTest’s ability to detect combinations. But the newest version did not play measurably better than the older version in head-to-head competition against each other. The extra computation cost of the fail-high singular extensions wiped out the gain. We decided to keep the new extensions, figuring that they might have a positive effect against human players. At the very least, the extensions would make it impossible for the human players to guess whether the program would see a particularly deep combination.

几年后,托马斯决定放弃他原来的论文主题,对奇异扩展和我们使用的其他一些搜索扩展进行详细研究。团队中存在一些争论,是否应该相信托马斯论文中的一些结论的表面意义。托马斯的结果是通过我们的下一代国际象棋机器 Deep Thought 使用未经证实的间接测量技术获得的。他的结果具有启发性,但它们提供的问题多于答案。Thomas 认为,单一扩展对程序性能影响不大,至少在 Deep Thought 中实现是这样。直到 1991 年,当 Deep Thought II 开始使用奇异扩展的全新实现时,我们才验证了奇异扩展的影响是相当大的。但这是在我们引入了一些额外的新想法之后,这些新想法在《Deep Thought》中没有出现,涉及如何控制奇异扩展及其导数。特别是,Deep Thought II 的实现还扩展了对并非真正单一的走法的搜索。例如,当这一步棋是仅有的两个好棋步之一时,深度思考 II 仍然会扩大搜索范围,尽管不会像该棋步是单一棋步时那么大。

A couple of years later, Thomas decided to abandon his original thesis topic and do a detailed study of singular extensions and a few other search extensions that we used. There was some debate among the team whether some of Thomas’ conclusions in his thesis should be taken at face value. Thomas’ results were obtained using an unproven indirect measurement technique with Deep Thought, our next chess machine. His results were suggestive, but they provided more questions than answers. According to Thomas, singular extensions had little effect on program performance, at least as implemented in Deep Thought. It was not until 1991, when Deep Thought II became operational with an entirely new implementation of singular extensions, that we verified that the effects of singular extensions were quite substantial. But this was after we introduced several additional new ideas, which were not present in Deep Thought, on how to control the singular extensions and their derivatives. In particular, the Deep Thought II implementation also extended the search on moves that were not really singular. For instance, when the move was one of the only two good moves, Deep Thought II would still extend the search, although not as much as when the move was singular.

击败“世界冠军”

Beating “World Champions”

在引入单一扩展后,ChipTest 与 Hitech 处于大致平等的地位。它在战术上比Hitech更强,但它的国际象棋知识几乎为零。我们本可以添加一些简单的软件评估来提高 ChipTest 的实力,但我有别的想法。

ChipTest was playing on roughly equal footing with Hitech after the introduction of singular extensions. It was tactically stronger than Hitech, but it had close to zero chess knowledge. We could have added some simple software evaluation to improve ChipTest’s strength, but I had something else in mind.

当我开始这个项目时,我计划设计一个评估芯片来配合移动生成器芯片。我对设计的外观有一些粗略的想法,但我不知道要加入什么样的国际象棋评估功能。有了 Hitech 团队的帮助,我希望能够借鉴当地的专业知识和经验。问题是 Hitech 的设计理念不同,与我将整个评估功能安装到单个芯片上的想法不兼容。当 ChipTest 启动并运行时,我得出的结论是,我需要先获得一些国际象棋评估功能的第一手设计经验,然后才能认真考虑评估芯片的想法。

When I started the project, I planned to design an evaluation chip to go with the move generator chip. I had some rough idea on how the design would look, but I had no idea on what kind of chess evaluation function to put in. With the Hitech team around, I hoped that I could draw on the local expertise and experience. The problem was that Hitech was designed with a different philosophy, and was not compatible with my idea of fitting an entire evaluation function onto a single chip. By the time that ChipTest was up and running, I concluded that I needed to gain some design experience with chess evaluation function first hand before I could seriously entertain the idea of an evaluation chip.

那时,第一批现场可编程门阵列 (FPGA) 刚刚投入商用。这些 FPGA 允许使用现成的组件来实现中等复杂度的逻辑电路。它们似乎非常适合在真实国际象棋机器中执行实验性硬件评估功能的任务。孔教授有一些可自由支配的资金,并同意支付费用。这台新机器主要是关于如何设计国际象棋评估硬件的学习体验。但是,根据 ChipTest 的性能,新机器有望成为比任何现有国际象棋机器都更强的棋手。我还希望在 1987 年的 ACM 锦标赛上大放异彩!然而,考虑到当时 FPGA 设计软件的粗糙性,以及我从未使用 FPGA 进行设计的事实,在 1987 年的锦标赛中进入新机器的机会并不好。

By then, the first field-programmable gate arrays (FPGAs) had just become commercially available. These FPGAs allowed logic circuits of moderate complexity to be implemented with off-the-shelf components. They seemed to be ideally suited for the task of an experimental hardware evaluation function in a real chess machine. Professor Kung had some discretionary funds and agreed to foot the bill. This new machine would be mainly a learning experience on how to design chess evaluation hardware. But, based on the performance of ChipTest, the new machine could be expected to be a stronger chess player than any existing chess machine. I was also hoping to kick some butts at the 1987 ACM tournament! Given the crude nature of FPGA design software at the time, and the fact that I had never designed with FPGAs, the chance of entering the new machine in the 1987 tournament, however, was not good.

距离 1987 年 ACM 锦标赛还有一个月,很明显,绝对没有机会及时准备好新机器。我们制定了应急计划。首先,我们在 ChipTest 中添加了 1986 年计划的棋子结构软件评估。其次,我们通过针对 ChipTest 本身和 Hitech 7 进行认真的游戏测试,确保硬件和软件都处于良好状态。第三,我们“获得”了我们能找到的最快的工作站。我们并不拥有工作站——它是借来的通过我们与工程实验室工作人员的联系。官方称,该工作站在安装到其真正所有者的实验室之前正在接受测试。更快的工作站减少了软件开销,ChipTest 每秒能够搜索大约 400,000 个国际象棋位置,速度提高了 30%。

One month before the 1987 ACM tournament, it became clear that there was absolutely no chance of getting the new machine ready in time. We drew up a contingency plan. First, we added to ChipTest the software evaluation of pawn structure that we had planned back in 1986. Second, we made sure that both the hardware and the software were in good shape by doing serious play testing, both against ChipTest itself and against Hitech7. Third, we “acquired” the fastest workstation that we could lay our hands on. We did not own the workstation—it was borrowed through our connections with the Engineering Lab staff. Officially, the workstation was being tested before its installation in its real owner’s lab. The faster workstation reduced the software overhead and ChipTest was able to search about 400,000 chess positions per second, a thirty percent increase in speed.

添加软件棋子结构评估后,ChipTest 在对战游戏中的评分提高了约 100 个点。由于穆雷正式进入 Hitech 组,所以一个大问题是 Hitech 是否会参加 ACM 锦标赛。它没有参加 1986 年的比赛,因为它的硬件正在修改。1987年,Berliner博士决定再次跳过比赛,理由是希望专注于与人类玩家的比赛,因此Murray继续从事ChipTest工作没有任何问题。

After the addition of the software pawn structure evaluation, ChipTest became about 100 rating points stronger in games against machines. Since Murray was officially in the Hitech group, one big question was whether Hitech would be playing in the ACM Championship. It did not play in the 1986 tournament because its hardware was being modified. Dr Berliner decided to skip the competition again in 1987, citing the desire to concentrate on competitions against human players, so there was no problem with Murray continuing to work on ChipTest.

尽管没有 Hitech,1987 年在达拉斯举行的 ACM 锦标赛仍然有强大的参赛队伍。1986年世界电脑象棋锦标赛的顶尖选手除了Hitech外全部都到场了。ChipTest 排名第三,落后于 1986 年世界计算机象棋锦标赛抢七赛冠军 Cray Blitz 和 1986 年 ACM 锦标赛冠军 Belle。贝贝和菲尼克斯分别是第四号和第五号种子,但在世界电脑象棋锦标赛上并列第一。

Despite the absence of Hitech, the 1987 ACM tournament in Dallas had a strong field. The top finishers in the 1986 World Computer Chess Championship were all present with the exception of Hitech. ChipTest was seeded third, behind Cray Blitz, the winner of the 1986 World Computer Chess Championship on tiebreak, and Belle, the winner of the 1986 ACM Championship. Bebe and Phoenix were seeded fourth and fifth respectively but tied for first at the World Computer Chess Championship.

ChipTest 在第一轮轻松击败了 Cyrus 68K。第二轮遇到了一个更难对付的对手,Lachex,它在 1986 年 ACM 锦标赛中获得第二名。ChipTest 在比赛一开始就赢得了棋子,但这是一场很难获胜的比赛,比赛一直持续到中部时间凌晨 3:00,并且是最后一个结束的。匹兹堡采用东部时间,所以从生物学上来说,对我来说现在是凌晨 4:00,远远超过了我的就寝时间。我很累,但很高兴赢得了比赛。

ChipTest beat Cyrus 68K in round one easily. Round two brought a more difficult opponent, Lachex, which had finished in second place in the 1986 ACM tournament. ChipTest won a pawn early in the game, but it was a difficult endgame to win, going on until 3:00 am, Central Time, and was the last to finish. Pittsburgh is in Eastern Time so, biologically, it was 4:00 am for me and way past my bedtime. I was very tired but happy to have won the game.

下一轮,我们登上了榜首,对阵 Cray Blitz。我们有黑色。

In the next round, we moved to the top board to play against Cray Blitz. We had Black.

ChipTest 的开局用国际象棋的术语来说就是中卫。辩护方的声誉并不好。对中卫的主要批评是,黑后在很早的阶段就移动到空旷的地方,可能会受到攻击。当我看到开口时,我皱起了眉头。我读过的一本国际象棋入门书给出了这一点以开局为例,说明如何在游戏的开局阶段不玩。穆雷没有太多时间准备开场,因为准备相对容易,所以决定采用防守。肯·汤普森过来问道:“你怎么能再玩这种垃圾游戏呢?” 在与Lachex的比赛中,ChipTest同样打的是中卫。看过拉切克斯的比赛后,克雷闪电战队为怀特准备了更好的进攻线。

ChipTest played an opening called the center counter defense, in chess parlance. The defense did not have a good reputation. The main criticism of the center counter defense is the fact that the Black queen moves out into the open at a very early stage and can be subject to attacks. I grimaced when I saw the opening. One of the introductory chess books that I had read gave this particular opening as an example of how not to play in the opening phase of the game. Murray did not have a lot of time to prepare the opening, and decided to use the defense since it was relatively easy to prepare. Ken Thompson came by and asked, “How can you play this garbage again?” In the Lachex game, ChipTest had also played the center counter defense. Having seen the Lachex game, the Cray Blitz team prepared a better offensive line for White.

开局不久后,ChipTest 的棋盘位置看起来有点危险,此时它突然预测 Cray Blitz 将赢得棋子。ChipTest 进入了我们所说的恐慌模式,试图通过所有动作来找到出路。经过大约三十分钟的计算,它进行了一次急剧的意外延续,仍然会失去棋子,但会得到一些补偿。赛事总监Mike Valvo对于任何程序在一步棋上花费三十分钟感到非常震惊,并称赞ChipTest在关键时刻明智地花费了时间。事实上,ChipTest 单次移动花费三十分钟的原因是,当与紧急模式结合使用时,单一扩展往往会导致搜索树的大小爆炸。我们在测试游戏中了解到了这一点,因此每当 ChipTest 处于恐慌模式时,我们都会给它分配很大的时间。

Pretty soon after the opening phase, the board position looked a little bit dicey for ChipTest, and at this point it suddenly predicted that Cray Blitz would win a pawn. ChipTest went into what we called panic mode, trying to find a way out by going through all the moves. After about thirty minutes of calculation, it went for a sharp surprise continuation that would still lose the pawn but with some compensation. Mike Valvo, the tournament director, was quite shocked that any program would spend thirty minutes on a single move, and praised ChipTest for spending time wisely at a critical juncture. In reality, the reason why ChipTest spent thirty minutes on a single move was because singular extensions tend to cause the search tree to explode in size when used in conjunction with the Panic Mode. We had learned about this in our test games and therefore we gave ChipTest a very large time allocation whenever it was in panic mode.

鉴于 ChipTest 的出人意料的举动,Cray Blitz 莫名其妙地决定不赢得棋子。如果 ChipTest 按照 Cray Blitz 的预测进行正常的延续,Cray Blitz 就会继续前进并立即抓住棋子。在 ChipTest 的整个 30 分钟计算过程中,Cray Blitz 一直在计算其对正常延续的响应。我松了一口气。克雷闪电战的开口边缘已经消散。这场比赛似乎只需要几步就可以了。但随后,ChipTest 突然宣布它获胜了。由于单一延伸,它计算出了一条很长的强迫线,并且报告说,在这条线的末端,它将赢得主要材料。克雷·布利茨根本不知道自己的地位即将像纸牌屋一样倒塌。克雷闪电战又走了三四步,才意识到比赛已经输了。这场比赛只持续了二十七步,可能是克雷·布利茨职业生涯迄今为止最短、最惨重的一场失败。这可能也是第一个单一扩展发挥主要作用的公共游戏。

Given the surprise move from ChipTest, Cray Blitz inexplicably decided not to win the pawn. Had ChipTest played the normal continuation that Cray Blitz was predicting, Cray Blitz would have gone ahead and grabbed the pawn immediately. Cray Blitz had been calculating its response to the normal continuation during the entire 30 minutes of ChipTest’s calculation. I breathed a sigh of relief. Cray Blitz’s opening edge had dissipated. The game seemed to be about even for a few moves. But then, out of nowhere, ChipTest declared that it was winning. It had calculated a very long forcing line as a result of singular extensions and, was reporting that at the end of the line, it would win major material. Cray Blitz simply had no idea that its position was about to collapse like a house of cards. It took Cray Blitz three or four more moves before it realized that the game was lost. The game lasted only twenty-seven moves and was probably the shortest and worst loss of Cray Blitz’s career up to that day. It was probably also the first public game where singular extensions played a major role.

上一轮,我们让白棋对阵凤凰,这是一个节目在工作站网络上运行。我们在开幕式上完全没有做好准备。Phoenix 在其开篇书中的举措比 ChipTest 多得多。由于 ChipTest 有白棋,开局准备不足并不是致命的。我们确实失去了先手优势,比赛进入了双方机会均等的残局。《凤凰城》的作者乔纳森·谢弗 (Jonathan Schaeffer) 是一位大师级国际象棋棋手;在柏林纳博士之后,乔纳森很可能是活跃的计算机国际象棋研究人员中最强的国际象棋棋手。前世界冠军米哈伊尔·博特维尼克(Mikhail Botvinnik)本来应该正在开发国际象棋程序,但他可能不能被视为活跃的计算机国际象棋研究员,因为他的团队之外没有人见过他的程序下过棋。作为一名实力较弱的玩家,我专心聆听 Jonathan 对这个职位的看法,并不断对照 ChipTest 对这个职位的评估来核实他的观点。乔纳森后来被引用说,我作为一个弱玩家,是那些情绪随着自己的程序对位置的评估而起伏的人之一。在这场比赛中确实如此。乔纳森(Jonathan)认为他的程序处于更好的位置,但由于 ChipTest 认为位置已经完全平等,因此我并不担心。当 ChipTest 的评价开始向有利于它的方向攀升时,乔纳森的心情仍然很愉快。然后乔纳森突然意识到菲尼克斯可能遇到了麻烦。凤凰的地位已经站不住脚了。ChipTest 已经找到了赢得材料的方法。几招之后,乔纳森(Jonathan)离开了凤凰城,ChipTest 横扫冠军。

In the last round, we had White against Phoenix, a program that ran on a network of workstations. We were thoroughly out-prepared in the opening. Phoenix had many more moves in its opening book than ChipTest. Since ChipTest had White, being out-prepared in the opening was not fatal. We did lose the advantage of the first move, and the game reached an endgame with equal possibilities for both sides. The author of Phoenix, Jonathan Schaeffer, was a master level chess player; after Dr Berliner, Jonathan might very well have been the strongest chess player among the active computer chess researchers. Former World Champion, Mikhail Botvinnik, was supposed to be working on a chess program, but he probably could not be considered an active computer chess researcher, as no one outside of his group had ever seen his program play a single move. Being a weak player myself, I listened intently to what Jonathan had to say about the position, and constantly checked his opinion against ChipTest’s assessment of the position. Jonathan was later quoted as saying that I, being a weak player, was one of those guys whose emotions went up and down with their own program’s evaluation of the position. This was certainly true in this game. Jonathan thought his program was in a better position, but with ChipTest thinking the position dead even, I was unconcerned. Jonathan was still in a cheerful mood when ChipTest’s evaluation started to creep upward in its favor. Then it dawned upon Jonathan that Phoenix might be in trouble. Phoenix’s position was no longer tenable. ChipTest had already seen ways to win material. A few moves later Jonathan resigned for Phoenix, and ChipTest won the Championship with a clean sweep.

ChipTest 可能是当时世界上最好的国际象棋机器它在 1986 年世界计算机国际象棋锦标赛中领先四名选手中的三名获得了 ACM 冠军。Hitech 是世界锦标赛前四名中唯一缺席的球队,而 ChipTest 在 ACM 锦标赛之前就曾与 Hitech 保持过联系。ChipTest有多强?我有理由相信,这可能是第一台高级大师机。ChipTest 从未在人类锦标赛中得到全面评级,因此这只是一个猜测。

ChipTest was probably at this point the best chess machine in the world. It had won the ACM title ahead of three of the top four finishers in the 1986 World Computer Chess Championship. Hitech was the only one among the top four at the World Championship that was missing, and ChipTest had held its own against Hitech before the ACM tournament. How strong was ChipTest? I have reasons to believe that it might have been the first senior master machine. ChipTest was never fully rated in human tournaments, so this is just a guess.

ChipTest 获胜后获得了 2000 美元的奖金,这比建造它的预计成本 500 美元还要高。为了庆祝胜利,我们用部分奖金为整个部门举办了一场 TG 派对,以纪念 ChipTest。考虑到 ChipTest 最初的诞生方式,这似乎非常合适。

ChipTest received a $2000 prize for the win, which was more than the estimated $500 cost of building it. To celebrate the win, we used part of the prize money to hold a TG party for the whole Department in ChipTest’s honor. It seemed very appropriate, given how ChipTest was born in the first place.

来自加州的邀请

An Invitation from California

在 1987 年 ACM 锦标赛的一次谈话中,《Phoenix》的作者乔纳森·谢弗 (Jonathan Schaeffer) 表示,在计算机国际象棋领域,他真正尊敬的有两个人。

During one of the conversations at the 1987 ACM Championship, Jonathan Schaeffer, the author of Phoenix, said that there were two people whom he really looked up to in the computer chess field.

第一个人是肯·汤普森。乔纳森尊重肯,不仅因为他的成就和对计算机国际象棋的贡献,还因为他的无私行为。他提到了肯的一些事迹,其中之一是我自己观察到的。每年在 ACM 锦标赛上,Ken 都会在每轮比赛结束后收集比赛成绩,将其输入计算机,然后将其发布到世界其他地方。你不会期望一个忙碌而重要的人获得图灵奖(相当于计算机科学家的诺贝尔奖)。

The first person was Ken Thompson. Jonathan respected Ken not just for his achievement and contribution to computer chess, but also for his unselfish acts. He mentioned a few of Ken’s deeds, one of which I observed myself. Each year at the ACM Championship, Ken would collect the game scores after each round, enter them into the computer, and then post them to the rest of the world. Not something you would expect from a busy and important person who had won the Turing Award, the equivalent of the Nobel Prize for a computer scientist.

乔纳森尊敬的第二个人让我大吃一惊。这是汉斯·柏林纳博士。回想起来,我感到惊讶,部分原因是我与柏林博士的专业差异,部分原因是我认为他不是最容易共事的人,但我想我可以理解为什么乔纳森对他评价很高。当乔纳森看着柏林博士时,他一定看到了自己的一部分,或者至少看到了他想成为的人。他们都是非常优秀的国际象棋棋手,正在研究计算机国际象棋。乔纳森仍是一位后起之秀,而柏林纳博士则已颇有建树。Berliner 博士自 1970 年开始从事计算机国际象棋研究,当时他在 ACM 比赛中输入了第一个程序。早前获得世界函授国际象棋冠军称号后,柏林纳博士可能希望自己能够成为最终解决计算机国际象棋问题的人,创造出一台可以在比赛中击败世界国际象棋冠军的国际象棋机器。他在相当高的年纪回到研究生院,获得了博士学位,并成为卡内基梅隆大学的一名研究人员,该大学的计算机科学系在计算机国际象棋领域有着开创性的工作历史。他聚集了一批技能各异的研究生,带领他们创建了高科技。很少有人能与柏林纳博士的干劲相媲美。

The second person that Jonathan looked up to came as a big surprise to me. It was Dr Hans Berliner. Looking back, I was surprised partly because of my professional differences with Dr Berliner, and partly because I considered him not the easiest person to work with, but I guess I can understand why Jonathan thought highly of him. When Jonathan looked at Dr Berliner, he must have seen quite a bit of himself, or at least something of what he would like to be. They were both very good chess players doing research in computer chess. Jonathan was still an up-and-comer, while Dr Berliner was well established. Dr Berliner had worked on computer chess since 1970 when he entered his first program in the ACM competition. Having garnered the title of World Correspondence Chess Champion earlier, Dr Berliner may have hoped that he could be the person who would finally solve the Computer Chess Problem, creating a chess machine that could beat the World Chess Champion in a match. He went back to graduate school at a fairly advanced age, earned a PhD degree, and became a research faculty member at Carnegie Mellon University, where the Computer Science Department had a history of pioneering work in computer chess. He gathered together a group of graduate students with varying skills and led them in the creation of Hitech. There are very few people who could match Dr Berliner’s drive.

在我与 Jonathan 交谈几周后,ChipTest 团队和 Berliner 博士之间的友好竞争突然变成了公开竞争。

A few weeks after my conversation with Jonathan, the friendly competition between the ChipTest team and Dr Berliner suddenly turned into an open rivalry.

在我告诉你这个故事之前,我需要把你放在正确的框架中的心灵。首先,故事很重要。这样的事情确实会发生。我不喜欢讲这个故事,但故事中的事件标志着 ChipTest 团队和 Berliner 博士之间关系的不可逆转的转变。其次,这个故事可能会让你对柏林博士产生不利的看法。在你这样做之前,请三思一下如果你代替他会做什么。想象一下,你有一个终生目标,一个更高的使命。为了实现这一目标,您已经工作了数十年,做出了许多个人牺牲。您现在是您所选择的领域中广泛认可的权威。过去几年你取得了巨大的成功,你的目标似乎就在眼前。现在想象一下一群学生不知从哪里冒出来。他们不是来自你的领域,他们有不同的想法。你对他们没有什么说服力,因为他们不是你的学生。他们的做法与您所提倡的完全相反。更糟糕的是,他们似乎正在快速进步,这使得说服他们遵循你的愿景变得更加困难。你会怎么办?你能做什么?这是一个非常艰难的情况。

Before I tell you the story, I need to put you in the right frame of mind. First, the story is important. Things like this do happen. I don’t enjoy telling it, but the events in the story marked an irreversible transition in the relationship between the ChipTest team and Dr Berliner. Second, the story may make you look at Dr Berliner unfavorably. Before you do that, please think twice about what you would have done in his place. Imagine that you have a lifetime goal, a higher calling. You have worked for decades to reach the goal, making many personal sacrifices. You are now a widely recognized authority in your chosen field. You have achieved great successes in the last few years and your goal seems to be in sight. Now imagine a bunch of students coming out of nowhere. They are not from your field and they have different ideas. You have little persuasive power over them since they are not your students. Their approach is the exact opposite of what you are advocating. To make matters worse, they seem to be making rapid progress, which makes it even harder to persuade them to follow your vision. What would you do? What can you do? It is a very tough situation.


我的同事安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 正在加利福尼亚州帕萨迪纳市的 NASA 喷气推进实验室 (JPL) 工作。他遇到了电脑象棋爱好者斯图尔特·克拉克拉夫特(Stuart Cracraft),当时他恰好在喷气推进实验室工作。Stuart 向 Andreas 询问了很多有关 ChipTest 的问题。一场大型国际象棋锦标赛美国公开赛即将在洛杉矶举行。Stuart 显然与锦标赛组织者有良好的联系,并建议我们让 ChipTest 参加锦标赛。斯图尔特提到,他有一些朋友很乐意帮助建立通信和远程操作机器。Andreas 认为这是一个很好的机会来看看 ChipTest 对抗人类玩家的表现如何。

Andreas Nowatzyk, my office mate, was spending time at NASA’s Jet Propulsion Lab (JPL) in Pasadena, California. He met Stuart Cracraft, a computer chess enthusiast, who happened to be working at JPL at the time. Stuart asked Andreas a lot of questions about ChipTest. A big chess tournament, the American Open, was about to take place in Los Angeles. Stuart apparently had good contacts with the tournament organizers and suggested that we enter ChipTest into the tournament. Stuart mentioned that he had friends who would gladly help with setting up communications and operating the machine remotely. Andreas thought that it was a good opportunity to see how well ChipTest would do against human players.

托马斯全力以赴参加比赛。穆雷不想选边站队,因为他和我都知道 Hitech 将参加同一场比赛。我反对参加,但并不是因为 Hitech 参加了比赛。我发现我花了太多时间看 ChipTest 比赛,却没有完成太多工作。我并没有像我希望的那样遵守纪律。如果 ChipTest 开始参加锦标赛,我正在开发的新机器可能永远不会面世。不管怎样,我给柏林博士发了一封电子邮件征求他的意见。

Thomas was all for entering the tournament. Murray did not want to take sides, as both he and I knew that Hitech was going to the same tournament. I was against participating, but not because Hitech was playing. I found that I was spending too much time watching ChipTest play and not getting much work done. I was not quite as disciplined as I would have liked. If ChipTest started playing in tournaments, the new machine that I was working on might never see the light of the day. Anyway, I sent an e-mail to Dr Berliner asking for his opinion.

第二天早上,Berliner 博士发回一条信息,强烈反对 ChipTest 继续进行,并给出了三​​个理由。首先,ChipTest 完全没有针对人类玩家进行过测试。第二是卡内基梅隆大学的声誉可能会受到损害。他的第三个观点是,让与大学无关的外人来运营 ChipTest 是愚蠢的。

The next morning, Dr Berliner sent back a message vehemently against ChipTest going, giving three reasons. The first was that ChipTest was totally untested against human players. The second was that Carnegie Mellon’s reputation could be hurt. His third was that it was foolish to let an outsider, not affiliated with the university, operate ChipTest.

在我看来,ChipTest 未经人类玩家测试是值得尝试的完美理由。毕竟我们是在一所研究型大学。此外,我从未听说过哪所大学的声誉因学生项目在国际象棋比赛中表现不佳而受到损害。Berliner 博士的第三个反对意见提出了一个合理的担忧,但 ChipTest 是一个零预算的学生项目,我没有看到任何强有力的理由反对获得外部帮助来操作机器。否则,我们就得花一大笔钱飞到洛杉矶8

From my point of view, ChipTest being untested against human players would have been the perfect reason to take it out for a spin. We were in a research university after all. Furthermore, I had never heard of a university’s reputation ever being ruined by student projects doing poorly in chess tournaments. Dr Berliner’s third objection brought up a valid concern, but ChipTest was a student project with a zero budget, and I did not see any strong reason against getting outside help to operate the machine. Otherwise, we would have to shell out a serious amount of our own money to fly to Los Angeles 8.

我不同意 Berliner 博士的推理,但由于我并不是真的想让 ChipTest 参加比赛,所以我只是将 Berliner 博士的消息转发给 Andreas。

I disagreed with Dr Berliner’s reasoning, but since I did not really want to enter ChipTest into the tournament I just forwarded Dr Berliner’s message to Andreas.

安德烈亚斯下午回复说,潜在的运营商确实与卡内基梅隆大学有联系。他就是吉姆·吉洛格利(Jim Gillogly),计算机国际象棋的早期先驱之一,与柏林纳几乎同时获得该系的博士学位。几个小时后,吉姆给我发了一封电子邮件,正式同意在需要时担任接线员。我把这个信息转达给了柏林博士。

Andreas replied in the afternoon that the potential operator did have an affiliation with CMU. He was Jim Gillogly, one of the early pioneers in computer chess and a PhD from the Department at roughly the same time as Berliner. A couple of hours later, Jim sent me an e-mail formally agreeing to be the operator, if needed. I passed this information on to Dr Berliner.

柏林纳博士可能措手不及。这次他想出了一个不同的理由。ChipTest 团队尚未发表有关奇异扩展的论文,但 Berliner 博士知道它们。Hitech 团队刚刚开始在他们的机器上进行单一扩展。由于 Hitech 没有 Singular Extensions,Berliner 博士担心它可能会落后于 ChipTest,而 Hitech 团队的表现会很糟糕。嗯,这也不是完全没有道理。我们自己提出了单一扩展的想法,没有得到柏林博士的任何帮助,所以从这个意义上说,他的要求是不合理的。另一方面,我们欠柏林博士一些东西。当 Murray 同意帮助 Thomas 和我开发 ChipTest 的评估功能时,他并不是从头开始。当时我不知道的是,穆雷使用了Berliner 博士编写的“Cray Blitz Simulator”代码9作为创建 ChipTest 评估函数的起始模板。Cray Blitz Simulator 是 Berliner 用来在 Hitech 上运行的一个简单的(与 Hitech 相比)评估函数,以模拟 Cray Blitz 的预期行为。Berliner博士的代码量还不到ChipTest总代码的百分之五,但它仍然是他的。我第一次了解到相关代码的本质是在 1987 年 ACM 锦标赛前几周。Hitech 小组的戈登·戈奇 (Gordon Goetsch) 开玩笑说克雷闪电战有可能被克雷闪电战模拟器打败。

Dr Berliner was probably taken off guard. This time he came up with a different reason. The ChipTest team had not published the paper on singular extensions yet, but Dr Berliner knew about them. The Hitech group was just in the process of getting singular extensions to work on their machine. Since Hitech did not have Singular Extensions, Dr Berliner worried that it could finish behind ChipTest, and the Hitech group would look bad. Well, it was not completely unreasonable. We had come up with the idea of singular extensions ourselves without any help from Dr Berliner, so in that sense his request was unreasonable. On the other hand, we owed Dr Berliner something. When Murray agreed to help Thomas and me with the evaluation function for ChipTest, he did not start from scratch. Unbeknown to me at the time, Murray used the “Cray Blitz Simulator” code 9 written by Dr Berliner as the starting template to create ChipTest’s evaluation function. The Cray Blitz Simulator was a simple (compared to Hitech’s) evaluation function used by Berliner to run on Hitech to simulate the expected behavior of Cray Blitz. The amount of code from Dr Berliner was less than a half of one percent of ChipTest’s total code, but it was his nonetheless. I first learned about the nature of the code in question only a few weeks before the 1987 ACM Championship. Gordon Goetsch, who was in the Hitech group, was joking about the possibility of Cray Blitz being outplayed by the Cray Blitz Simulator.

如果柏林纳博士一开始就提出这一反对意见,我们可能就会尊重他的要求。这时候我已经有点生他的气了。我和托马斯进行了讨论,我们达成了妥协。只有当 Hitech 的单一扩展工作正常时我们才会参加比赛。事实上,我们没有制定任何旅行计划,因为我们无法想象他们能让他们及时工作。我将这一决定告知了柏林人和安德烈亚斯。

If Dr Berliner had stated this objection right at the beginning, we probably would have just honored his request. By this time, I was a little bit mad at him. I had a discussion with Thomas and we came up with a compromise. We would go to the tournament only if Hitech has the singular extensions working. In reality, we were not making any travelling plans as we could not imagine they could have them working in time. I informed both Berliner and Andreas of the decision.

安德烈亚斯推测,柏林纳博士反对的另一个可能原因可能是因为 Hitech 即将获得重大奖项,而柏林纳不想在岗位上被击败——考虑到两支球队是潜在的竞争对手,这是完全可以理解的立场。早在 20 世纪 70 年代末,爱德华·弗雷德金 (Edward Fredkin) 教授就设立了 10 万美元的奖金,奖励第一台在比赛中击败世界国际象棋冠军的计算机。卡内基梅隆大学受托管理弗莱金奖,柏林纳博士担任弗莱金奖委员会主席。为了鼓励该领域的稳步进步,还设立了两个较小的奖项:弗雷德金中级奖。Belle 团队凭借大师级表演荣获首届 Fredkin 中级奖。也许柏林纳博士已经将目光投向了自 Hitech 成立以来的第二个弗莱德金中级大师级表演奖。不完全了解相关规则获奖后,Andreas 指出,凭借独特的扩展,Hitech 可能足以达到 2400 分,成为第一位计算机高级大师。然而,中级二等奖的规则规定了25场比赛的成绩要超过2500。我不认为Hitech会足够好,即使有单一的扩展。我也不相信 ChipTest 能做到。如果 Berliner 博士希望 Hitech 成为第一位正式的高级大师,那么祝他好运。我还有更大的鱼要煎。无论如何,我们同意我们不会去,尽管有一个附带条件。

Andreas surmised that another possible reason for Dr Berliner’s objection, may have been because Hitech was within sight of a major prize and Berliner did not want to be beaten at the post—a perfectly understandable stance given that the two teams were potential rivals. Back in the late 1970s, Professor Edward Fredkin set up a $100,000 prize for the first computer to defeat the World Chess Champion in a match. Carnegie Mellon University was entrusted with managing the Fredkin Prize, and Dr Berliner was the chairman of the Fredkin Prize Committee. To encourage steady progress in the field, two additional but smaller prizes, the Fredkin Intermediate Prizes, were set up. The Belle team had claimed the first Fredkin Intermediate Prize for master level performance. Perhaps Dr Berliner had set his eyes on the second Fredkin Intermediate Prize for Grandmaster level performance ever since Hitech had been built. Not fully aware of the rules for the award, Andreas pointed out that, with singular extensions, Hitech might just be good enough to achieve a 2400 rating and become the first computer senior master. However, the rules for the second Intermediate Prize specified a twenty-five-game performance of over 2500. I did not believe Hitech would be good enough, even with the singular extensions. Neither did I believe that ChipTest could do it. If Dr Berliner wanted Hitech to be the first official senior master then good luck to him. I had bigger fish to fry. Anyway, we had agreed that we would not go, although with a proviso.

对于 Berliner 博士来说,Hitech 恰好是他的最后一台国际象棋机器——它花了数年时间才完成,并在之后的十多年里一直使用。ChipTest 花了七周的时间才完成,当时只有一年。但 ChipTest 实际上已完全折旧,获得的奖金比其成本还要多,而且我正在准备一台新机器。对我来说,ChipTest 错过一场比赛并不像完成新机器那么重要。

For Dr Berliner, Hitech happened to be his last chess machine—it took years to complete and stayed in service for over a decade afterwards. ChipTest took seven weeks to put together, and was only one year old at the time. But ChipTest was effectively fully depreciated, having earned more in prize money than it had cost, and I had a new machine in the pipeline. ChipTest missing a tournament was not as important to me as getting the new machine finished.

不久之后,孔教授召开会议讨论此事,我和柏林博士都参加了。穆雷和托马斯均未出席。会议的第一部分是谈论来自加州的邀请。Berliner博士首先发言,表达了他对ChipTest使用“Cray Blitz Simulator”代码的不满,这让我感到惊讶。直到 1986 年 ACM 锦标赛之后我才见过这个代码,但当我知道它的起源后,我认为穆雷在柏林博士的祝福下使用了它。在前一年的大部分时间里,我一直认为穆雷自己编写了代码。如果 Berliner 博士告诉我们他的不满,我们就会删除代码并从头开始重写。正如之前所指出的,无论如何,它还不到总代码的百分之二分之一。我解释说我对此一无所知,而默里才是对代码一无所知的人。(我当时并不知道,但托马斯可能是检索到密码的人。)柏林纳博士接受了我的解释。然后我重申我们不想参加比赛,所以邀请的问题终于结束了。

Shortly afterwards, Professor Kung called for a meeting to discuss things and both Dr Berliner and I attended. Murray and Thomas were not present. The first part of the meeting was to talk about the invitation from California. Dr Berliner made his statement first, expressing his dissatisfaction that ChipTest was using the “Cray Blitz Simulator” code which was a surprise to me. I had never seen the code until after the 1986 ACM Championship but once I knew its origin I thought that Murray used it with Dr Berliner’s blessing. For most of the year before, I had assumed that Murray wrote the code himself. Had Dr Berliner told us about his displeasure, we would have gutted the code and rewritten things from scratch. As it was previously noted, it was less than a half of one percent of the total code anyway. I explained that I knew nothing about it, and Murray was the one who knew anything about the code. (I did not know it at the time, but Thomas might have been the one who retrieved the code.) Dr Berliner accepted my explanation. I then reiterated that we did not want to play in the tournament, so the problem with the invitation was finally over.

会议的第二部分是关于我的未来。系里一直在考虑我毕业后留用我的可能性。当然,他们更希望柏林博士和我一起工作。我确实很享受在部门的生活,但我有一个主要的担忧。那信用呢?柏林博士回答“信用会自行解决”。我不太清楚他的意思,但我已经看到了 Hitech 发生的事情。大多数人将其称为柏林博士的机器,卡尔·埃贝林在很大程度上不为外界所知。这种看法很可能是由于媒体的不准确报道造成的,但这种假设仍然存在,我不希望在我从事的任何项目上发生类似的事情。嗯,这是一个有趣的提议,但我不得不拒绝,因为我不相信信用确实会自行解决。毕业后,我必须到别处寻找工作才能继续工作。当时计算机科学博士的就业前景并不好。哦,好吧,该怎样就怎样吧。我们要做的事情是让新机器尽可能成功,特别是争取弗雷德金中级奖。我们还需要让所有人都明白,新机器与 Hitech 团队的工作完全无关。

The second part of the meeting was about my future. The Department had been mulling over the possibility of retaining me after my graduation. Naturally, they would prefer that Dr Berliner and I work together. I did enjoy life a great deal in the Department, but I had one major concern. What about the credit? Dr Berliner answered “The credit will take care of itself”. I was not sure exactly what he meant by this, but I had seen what happened with Hitech. Most people referred to it as Dr Berliner’s machine and Carl Ebeling was largely unknown to the outside world. This perception may well have been due to inaccurate reporting in the press, but the assumption stuck and I would not want anything similar to happen on any project I worked on. Well, it was an interesting offer, but I had to decline, since I had no confidence that the credit would indeed take care of itself. After graduation I would have to look for a job elsewhere to continue the work. The job prospects for a computer science PhD were not good at the time. Oh, well, what will be will be. The thing to do would be to make the new machine as successful as we could, in particular, to go after the Fredkin Intermediate Prize. We also needed to make it obvious to anyone that the new machine was completely separate from the work of the Hitech team.

经历这件事后,我对信用问题变得更加敏感。我不想被虐待,也不想看到别人被虐待。会后我发誓我们的团队不会有正式的领导者。团队按职能划分,任命某个人为领导者对任何人都是不公平的。

I became a lot more sensitive to the credit issue after this episode. I didn’t want to be mistreated, and I didn’t want to see people mistreated. I swore after the meeting that our team would not have an official leader. The team was functionally partitioned, and it would be unfair to any one to declare a particular person as the leader.


回想起来,我想我是否只是幸运。到目前为止,我所有的项目经验都是某种合作伙伴关系——我在台湾的本科项目和通用电气项目都是与合作伙伴合作的。我的教职员老板的态度是不干涉,让我们做任何我们想做的事。因此,没有正式领导人的决定是很自然的。事实证明,国际象棋项目是一个需要团队每个成员完全奉献的项目。这种奉献之所以可能,只是因为我们是平等的伙伴。

Looking back, I wonder whether I was just lucky. All my project experience up to that point had been some sort of partnership—my undergradate project in Taiwan and the GE project had both been with partners. The attitude of my faculty bosses had been hands off and just letting us do whatever we pleased. So the decision to have no official leader was kind of natural. The chess project turned out to be a project that required complete dedication from every member of the team. This dedication was possible only because we were equal partners.

有几本书和文章指出,Hitech 团队和 ChipTest/Deep Thought 团队是死敌。这种关系比这更加文明。对于 Deep Thought 团队来说,这很容易——我们最终成为了胜利者。而且,在加州的邀请之后,事情才真正变得糟糕。可以肯定地说,事后双方都不会邀请对方参加他们的私人聚会。离开卡内基梅隆大学后,时间和距离帮助我们治愈了伤口;我认为双方都缓和了。当深思想II变得活跃,我们不再将Hitech视为竞争对手。我们仍然不喜欢输给 Hitech。天啊,我们不想输给任何人,包括加里·卡斯帕罗夫!

Several books and articles stated that the Hitech team and the ChipTest/Deep Thought team were bitter enemies. The relationship was more civil than that. For the Deep Thought team, it was easy—we ended up on the winning side. Also, things only really turned bad after the invitation from California. It is safe to say that neither team would invite the other to their private parties afterwards. After we left Carnegie Mellon, time and distance helped to heal the wounds; I think both sides mellowed. When Deep Thought II became active, we no longer considered Hitech as a competitor. We still did not like the idea of losing to Hitech. Hell, we did not like to lose to anybody, Garry Kasparov included!

    后来,当我们的国际象棋机器 ChipTest 和 Deep Thought 取得重大锦标赛胜利时,我们自己赞助了四台 TG 来庆祝胜利。

    Later on when our chess machines, ChipTest and Deep Thought, scored major tournament victories, we sponsored four TGs ourselves to celebrate the wins.

    国际象棋程序在搜索过程中遍历的位置可以被视为一棵搜索树,其中物理棋盘上的位置作为根位置。搜索到的棋步成为树枝,搜索到的位置成为分支点。

    The positions traversed by a chess program during a search can be viewed as a search tree, with the position on the physical chessboard acting as the root position. The moves searched become the tree branches, and the positions searched become the branching points.

    托尼可能是第一个使用这句话的人。

    Tony might have been the first person to use this phrase.

    、棋手从上往下依次为:世界冠军、特级大师(GM)、国际大师(IM)、高级大师(SM)、大师、专家、A级、B级、C级等在。前三个是国际冠军。FIDE大师(FM)是一个相对较新的国际头衔。FM的演奏强度大致相当于高级大师。超级宗师是非常强的宗师,但这个词主要是一种褒义,并不是一个正式的称号。

    From the top, the chess players can be ranked in the following order: World Champion, Grandmaster (GM), International Master (IM), senior master (SM), master, expert, class A, class B, class C and so on. The first three are international titles. FIDE Master (FM) is a relatively new international title. An FM is roughly comparable to a senior master in playing strength. A Super Grandmaster is a very strong Grandmaster, but the term is mainly one of praise and not an official title.

    这种安排无意中让公众产生了 ChipTest 团队为 Berliner 博士工作的印象,这后来成为摩擦的根源。

    The arrangement created an unintentional public perception that the ChipTest team worked for Dr Berliner, which later became a source of friction.

    Berliner 博士可以不受限制地访问 ChipTest,并广泛将其用作 Hitech 的陪练伙伴。我们保留了所有 ChipTest 游戏的日志,因此 Berliner 博士玩的陪练游戏对我们也很有用。我们还与 Hitech 进行了游戏来调试 ChipTest。Murray 也曾在 Hitech 工作,他为 ChipTest 团队进行了调试。

    Dr Berliner had unrestricted access to ChipTest which he used as a sparring partner for Hitech extensively. We kept logs of all ChipTest games, so the sparring games played by Dr Berliner were of use to us as well. We also ran games against Hitech to debug ChipTest. Murray, who also worked on Hitech, conducted the debugging sessions for the ChipTest team.

    我们自己并没有对 Hitech 进行太多的游戏测试。Berliner 博士使用 ChipTest 作为对抗 Hitech 的陪练伙伴。我们刚刚检查了游戏日志以确保没有发生任何奇怪的事情。

    We did not really do much play testing against Hitech ourselves. Dr Berliner was using ChipTest as a sparring partner against Hitech. We just checked out the game logs to make sure that nothing strange happened.

    该系确实支付了参加 ACM 锦标赛的费用,但参加 ACM 锦标赛是一项学术活动。

    The Department did pay for the trips to the ACM Championship, but competing in the ACM Championship was an academic activity.

    ChipTest 最终如何使用 Berliner 博士的这段代码?我们谁都不能确定。托马斯·阿南塔拉曼 (Thomas Anantharaman) 相信,在戈登·戈奇 (Gordon Goetsch) 向他建议后,他可能已经检索到了该代码。反过来,戈登可能只是想帮助我们在短短的七周内获得 1986 年 ACM 冠军。然而,戈登无法证实或否认托马斯的说法。

    How did ChipTest end up using this code from Dr Berliner? None of us knows for sure any more. Thomas Anantharaman believes that he might have retrieved the code after Gordon Goetsch suggested it to him. Gordon, in turn, might have been simply trying to help us make it to the 1986 ACM Champion in the short seven weeks that we had. Gordon, however, could not confirm nor deny what Thomas said.

 

 

第 5 章

CHAPTER 5

第一机器大师之争

The Race for First Machine Grandmaster


大约2200年代

Some 2200s

20 世纪 70 年代和 80 年代,英国科幻作家道格拉斯·亚当斯 (Douglas Adams) 创作了广受欢迎的科幻三部曲,BBC 将其改编成广播节目和电视剧。当我在台湾陆军服役时,当地一家广播电台正在播放根据第一本书《银河系漫游指南》改编的 BBC 广播节目。然而,直到我住在美国之后,我才开始阅读这些书。

During the 1970s and 1980s, the British science fiction writer Douglas Adams wrote a highly popular science fiction trilogy that the BBC turned into both a radio show and a television series. While I was serving in the Taiwanese Army, a local radio station was carrying the BBC radio show based on the first book, The Hitchhiker’s Guide to the Galaxy. However, I did not get to read the books themselves until I lived in the United States.

在《搭便车指南》的宇宙中,一个外星种族建造了一台名为“深思”的计算机,以解决“生命、宇宙和一切”的问题。经过亿万次计算,它得出了答案“42”。但包括深思在内的任何实体都不再知道最初的问题了。然而,Deep Thought 作为有史以来最强大的计算机,知道如何构建更强大的计算机来回答问题,问题是什么,答案“42”。

In the universe of the Hitchhiker’s Guide, an alien race built a computer called Deep Thought, to solve the question of “Life, the Universe, and Everything”. After eons of computation, it came up with the answer “42”. But no entity, including Deep Thought, knew the original question any more. Deep Thought, however, being the most powerful computer ever built, knew how to build an even more powerful computer which would answer the question, of what was the question, to the answer “42”.

嗯,能够击败世界冠军的机器,绝对名副其实的“深思”。

Well, a machine that could defeat the World Champion would surely be worthy of the name Deep Thought.

我不相信我们在 1988 年初制造的机器,也是我们的第一台真正的国际象棋机器,足以击败世界冠军。它可能至少慢了 100 倍。考虑到速度差异,新程序的第一个版本被称为 Deep Thought 0.01。在我们获得双处理器版本后,它后来成为 Deep Thought 0.02。最终,当我们决定加入 IBM 并使用新名称构建一台新机器时,该名称就变成了 Deep Thought。

I did not believe that the machine we were building in early 1988, which was also our first real chess machine, would be sufficient to defeat the World Champion. It was probably at least 100 times too slow. Factoring in the speed discrepancy, the first version of the new program was called Deep Thought 0.01. It later became Deep Thought 0.02 after we got a dual-processor version working. Eventually, when we decided to join IBM and build a new machine with a new name, the name became simply Deep Thought.

安德烈亚斯新年后从帕萨迪纳回到卡内基梅隆大学。我们断断续续地谈论创造的可能性一个可以自动调整国际象棋评估函数的程序。安德烈亚斯有一个令人兴奋的新想法。然而,新想法要求以不同的方式编写评估函数。Murray 喜欢 Andreas 的想法,并同意按照 Andreas 指定的方式编写一个全新的评估函数。新的评估函数的编写部分是为了处理即将完成的 Deep Thought 硬件支持的新评估原语。然而,新评估功能和自动调整软件的最初目的是从头开始。我们需要避免“Cray Blitz Simulator”代码中出现的任何问题,就像我们在 ChipTest 中遇到的那样,这样我们就能够自由地竞争。评估函数的自动调整经常被认为是 Deep Thought 中最新颖的方面,但这纯粹是出于必要。Andreas 在早期提出 ChipTest 方面提供了相当大的帮助,随着这种参与的增加,我终于让他同意被视为 Deep Thought 的作者之一。

Andreas came back to Carnegie Mellon from Pasadena after the New Year. We talked off and on about the possibility of creating a program that could tune the chess evaluation function automatically. Andreas had an exciting new idea. The new idea, however, required that the evaluation function be written in a different way. Murray liked Andreas’ idea and agreed to write a completely new evaluation function in the way Andreas specified. The new evaluation function was written in part to handle the new evaluation primitives supported by the Deep Thought hardware, which were about to be completed. The original reason for the new evaluation function, and the automatic tuning software, however, was to start from a clean slate. We needed to avoid any of the problems with the “Cray Blitz Simulator” code as we had had with ChipTest, so that we would be able to compete freely. The automatic tuning of the evaluation function was often mentioned as the most novel aspect of Deep Thought, but it came purely out of necessity. Andreas helped quite a bit in bringing up ChipTest earlier, and with this increased involvement I finally got him to agree to be seen as one of the authors of Deep Thought.

当安德烈亚斯认真参与后,我把目光投向了我的另一位同事迈克·布朗。(好吧,成为我的朋友有它的缺点。) ChipTest 最大的问题之一是它的开本,或者更确切地说,它缺乏开本。迈克同意编写一个程序,自动检查所有开局线,以寻找开局陷阱(糟糕的开局动作)和理论新颖性(有趣的新开局动作),并成为 Deep Thought 团队的第五位成员。

Once Andreas became seriously involved, I set my sight on Mike Browne, my other office mate. (Okay, being my friend has its drawbacks.) One of the biggest problems for ChipTest was its opening book, or rather, its lack of opening book. Mike agreed to write a program that automatically examined all the opening lines to look for opening traps (bad opening moves), and theoretical novelties (new opening moves that are interesting), and became the fifth member of the Deep Thought team.

在 20 世纪 80 年代的大部分时间里,弗雷德金基金会赞助了一系列年度计算机国际象棋赛事。该系列赛的最后一场比赛于 1988 年举行,即 ChipTest 赢得 ACM 冠军后的第二年。用于赞助活动的资金来自 10 万美元的弗莱德金奖奖金所产生的利息。Fredkin 赛事与 ACM 锦标赛中的计算机比赛不同,它让顶级国际象棋程序与人类国际象棋棋手进行对抗。传统上,ACM 计算机国际象棋卫冕冠军以及当时可能一些更强大的程序都会被邀请。1988 年的决赛是 Fredkin Masters Open,于 5 月 28 日至 30 日在卡内基梅隆大学校园举行。这将是一场六轮比赛,每天两轮。ChipTest 作为 ACM 卫冕冠军,与 Deep Thought 0.01 一起受邀参加。Hitech作为1985年ACM冠军也受邀参加。Hitech 将首次在锦标赛中使用 Singular Extensions。

During most of the 1980s, the Fredkin Foundation sponsored a series of annual computer chess events. The final event of the series took place in 1988, the year after ChipTest’s win at the ACM championship. The money used to sponsor the events came from the interest accrued from the prize fund for the $100,000 Fredkin Prize. The Fredkin events, unlike the computer competition in the ACM championship, pitted the top chess programs against human chess players. Traditionally, the reigning ACM Computer Chess Champion was invited along with possibly some of the stronger programs at the time. The 1988 final edition was the Fredkin Masters Open, from May 28 to 30 on the Carnegie Mellon campus. It would be a six-round tournament, with two rounds per day. ChipTest, as the reigning ACM Champion, was invited, along with Deep Thought 0.01. Hitech, as the 1985 ACM Champion, was also invited. Hitech would be using Singular Extensions for the first time in tournament play.

《深思1》的两个版本。

The two versions of Deep Thought 1.

Deep Thought 1 板的绕线(顶部)和印刷电路(底部)版本。绕线板右侧的两个线环用于将两个关键信号延迟几纳秒。(电信号每纳秒在电线上传播略小于一英尺。)

The wire-wrapped (top) and the printed circuit (bottom) versions of the Deep Thought 1 board. The two wire loops on the right side of the wire-wrapped board are used to delay two critical signals by a few nanoseconds. (Electrical signals travel slightly less than one foot over the wires every nanosecond.)

当 Deep Thought 硬件的工作开始时,我希望它能为 1987 年 ACM 锦标赛做好准备,但 Deep Thought 从未成功。我不会错过弗莱德金大师公开赛;我不会错过。然而,我最终把它剪得非常接近。与 ChipTest 相比,Deep Thought 中最重要的新功能是其硬件评估功能,它可以识别许多动态位置特征,包括棋子结构、通过的棋子、开档上的车等国际象棋概念。新的硬件评估是通过现场可编程门阵列实现的,其编程时间比我预期的要长。绕线的 Deep Thought 板也存在一些令人讨厌的电气问题,需要一段时间才能找到。而且,Deep Thought板除了具有新的硬件评估功能外,还有两个国际象棋处理器,都需要进行充分的测试。由于复杂性增加,Deep Thought 板的成本接近 5000 美元,是 ChipTest 成本的十倍。仅仅为 Deep Thought 采购零件就比设计和构建 ChipTest 的时间还要长。到 1988 年 3 月,Deep Thought 板已部分工作,但板的调试以及新软件的创建似乎需要很长时间。弗雷德金大师公开赛前两天,Deep Thought 板仍在​​重新布线,托马斯继续对新软件进行最后的修饰,直到赛事开始。

When the work began on the Deep Thought hardware, I was hoping to get it ready for the 1987 ACM Championship but Deep Thought never made it. I was not going to miss the Fredkin Masters Open; however, I ended up cutting it very close. The most important new feature in Deep Thought, compared to ChipTest, was its hardware evaluation function which recognized many dynamic positional features, including chess concepts such as pawn structure, passed pawns, rooks on open files, and so on. The new hardware evaluation was implemented with field programmable gate arrays that took longer to program than I expected. The wire-wrapped Deep Thought board also had some nasty electrical problems that took a while to track down. Furthermore, the Deep Thought board, besides having the new hardware evaluation function, had two chess processors, and both needed to be fully tested. With the additional complexity, the cost for the Deep Thought board was close to $5000, or ten times the cost of ChipTest. Just procuring the parts for Deep Thought took longer than the time to design and build ChipTest. By March 1988, the Deep Thought board was partially working, but the debugging of the board, along with the creation of the new software, seemed to take forever. Two days before the Fredkin Masters Open, the Deep Thought board was still being rewired, and Thomas continued to add finishing touches to the new software until the event started.

匹兹堡国际象棋俱乐部主办了这场比赛。参与的玩家约有三十人,其中大师级以上的约有二十人。其中一位外地大师是亚历山大·伊万诺夫(Alexander Ivanov),他是一位新近的苏联移民。亚历山大在美国评级量表上的评级为2597,这接近宗师实力,尽管他当时没有头衔。(大约一年后,他成为特级大师。) 匹兹堡顶级大师是维韦克·拉奥 (Vivek Rao),一位评分为 2491 的青少年,同时也是宾夕法尼亚州的顶级棋手。Vivek 的评级使他接近国际大师 (IM) 的实力。来自匹兹堡的大师被认为是对计算机特别挑剔的人群。他们中的许多人曾参加过早期版本的 Fredkin 赛事或玩过 Hitech。对于计算机来说,这将是一个非常好的测试。

The Pittsburgh Chess Club ran the tournament. About thirty players participated, approximately twenty at the master level or above. One of the out-of-town masters was Alexander Ivanov, a recent Soviet émigré. Alexander was rated 2597 on the US rating scale, which was close to Grandmaster strength, even though he did not have a title at the time. (He became a Grandmaster about a year later.) The top Pittsburgh master was Vivek Rao, a teenager rated at 2491 who also happened to be the top player in Pennsylvania. Vivek’s rating put him at close to International Master (IM) strength. The masters from Pittsburgh were thought to be a particularly tough crowd for computers. Many of them had participated in earlier editions of the Fredkin events or played Hitech before. It would be a very good test for the computers.

我使用 ChipTest 从 ACM 锦标赛获得的奖金支付了 ChipTest 和 Deep Thought 的比赛报名费和美国国际象棋联合会 (USCF) 会员费。付款后,我查看了参加者的挂图。当时 ChipTest 和 Deep Thought 都没有被评级,我很惊讶查看他们名字旁边的评级。我问了比赛总监,他说两台机器的评级都是柏林纳博士的估计。这两个项目都列为2200,本质上是非常弱的硕士。Hitech当时的评级为2376。我皱起了眉头,但没有说什么,因为我可以理解这些评级的原因。强项中的低预估评级意味着对于 ChipTest 和 Deep Thought 来说,我们的第一轮对手将是强大的大师。

I paid the tournament entry fees and US Chess Federation (USCF) membership dues for both ChipTest and Deep Thought, using ChipTest’s prize money from the ACM Championship. After paying, I checked the wall chart for the participants. Neither ChipTest nor Deep Thought was rated at the time, and I was surprised to see ratings placed next to the their names. I asked the tournament director, and he said that the ratings for both machines were Dr Berliner’s estimates. The programs were both listed as 2200, essentially very weak masters. Hitech was rated 2376 at the time. I winced but did not say anything as I could understand the reasoning for these ratings. The low estimate ratings in the strong field meant that for both ChipTest and Deep Thought our first round opponents would be strong masters.

Thomas 是 ChipTest 的操作员,我操作 Deep Thought。Deep Thought的第一个对手Ross Sprague,挂图评分为2339,是一位来自外地的律师,打法保守。到了第二天早上,比赛还没有分出胜负。与此同时,ChipTest 首次战胜了人类对手,一位强大的大师。

Thomas was the operator for ChipTest and I operated Deep Thought. Deep Thought’s first opponent Ross Sprague, rated 2339 on the wall chart, was a lawyer from out of town who played conservatively. The game was still undecided when it was time to adjourn till the next morning. Meanwhile, ChipTest scored its first win against a human opponent, a strong master.

第二天我很早就回来和罗斯一起完成比赛。我们都知道比赛即将平局,但罗斯可以采取冒险的举动来避免平局。但首先,他玩了一段时间,看看“深思”是否会自毁。最后,他认为评分为 2339 的人类玩家应该击败评分为 2200 的机器,并决定冒险。深思的评价顿时蹿升,几步之后,罗兹就认输了。如果不是 2200 的收视率估计,Deep Thought 可能会平局。第二轮比赛中,Deep Thought和ChipTest均战平强手。

I came back early the next day to finish the game with Ross. We both knew that the game was heading to a draw but that Ross could try to avoid the draw by playing a risky move. But first, he played on for a while to see whether Deep Thought would self-destruct. Finally, he figured that a 2339-rated human player should beat a 2200 machine and decided to take the risk. Deep Thought’s evaluation immediately jumped up, and a few moves later, Ross threw in the towel. Deep Thought probably would have drawn the game, if not for the 2200 rating estimate. In the second round games, both Deep Thought and ChipTest drew against strong masters.

Deep Thought第三轮的对手是金博尔·内德维德。金博尔的儿子鲁迪在我们部门工作,金博尔本人对国际象棋计算机非常熟悉。他之前曾输给海泰,也曾输过。金博尔非常自在,在比赛过程中他会时不时地发表评论。他在上一轮中与ChipTest打平,并对ChipTest的表现表示赞赏。深思还不太稳定,突然因为内部的不一致性而瞬间打出了一招。当我下这个动作时,我扬起了眉毛。金博尔不知道我为什么这么做,疑惑地说:“为什么?这是一个很好的举措。” 好吧,这只是碰巧而已。游戏中没有再发生任何事件,金博尔辞职了,却没有意识到 Deep Thought 可能会再次遇到同样的问题。与此同时,ChipTest 在一场令人毛骨悚然的比赛中赢得了第三轮比赛。第四轮,Deep Thought 再赢一局,ChipTest 打平。我们将在第三天见到顶尖选手。

Deep Thought’s third round opponent was Kimball Nedved. Kimball’s son, Rudy, worked in our department, and Kimball himself was quite familiar with chess computers. He had both beaten and lost to Hitech before. Kimball was sufficiently comfortable that he would make comments from time to time during the game. He had drawn ChipTest in the previous round and was complimentary of ChipTest’s play. Deep Thought was not quite stable yet, and all of a sudden, it played an instant move due to an internal inconsistency. I raised my eyebrows when I played the move. Kimball had no idea why I did this, and said questioningly, “Why? That is a good move.” Well, it just happened to be. There were no more incidents in the game, and Kimball resigned without ever realizing that Deep Thought could run into the same problem again. Meanwhile, ChipTest won its third round in a hair-raising game. The fourth round saw Deep Thought winning another game and ChipTest drawing its game. We would be meeting the top players on the third day.

第五轮中,Deep Thought 对阵 Alexander Ivanov,ChipTest 对阵 Vivek Rao。Deep Thought 仍然使用 ChipTest 在 1987 年 ACM 锦标赛中使用的同一开场书,当然,这意味着我们看到的是中后卫,也称为斯堪的纳维亚防守。我很惊讶地看到即将成为大师的亚历山大在开幕式上花了很多时间。他赛前确实说过,他从来没有认真研究过斯堪的纳维亚的防守。或许中卫也不是垃圾。亚历山大打得谨慎而有力。然后他对一个对深思来说看起来相当危险的立场进行了深入的思考。在我看来,亚历山大可以对深思的国王发动非常强大的攻击。令我惊讶的是,亚历山大的这一举动似乎释放了这个位置上所有的紧张。他在做什么?对我来说,这看起来是一个均匀的结局。仔细观察主板,我开始有一种沉沦的感觉。这不是一个均匀的结局。亚历山大比“深思”看得更深刻,而“深思”却输掉了第一场比赛。当 Deep Thought 遇到问题时,ChipTest 的董事会上正在发生有趣的事情。在 Vivek 与 ChipTest 比赛之前,他公开表达了他对下棋计算机的蔑视,他在早期的遭遇中以牺牲 Hitech 为代价,获得了许多美好的回忆。ChipTest 在二十二步棋后迫使 Vivek 辞职,做出了意想不到的牺牲。比赛结束几个小时后,Vivek 仍在向其他人展示 ChipTest 的惊喜组合。

In the fifth round, Deep Thought was paired against Alexander Ivanov, and ChipTest against Vivek Rao. Deep Thought was still using the same opening book that ChipTest used in the 1987 ACM Championship, and that, of course, meant we were seeing the center counter defense, also known as the Scandinavian defense. I was surprised to see Alexander, the soon-to-be Grandmaster, spending a lot of time in the opening. He did say before the game that he had never studied the Scandinavian defense seriously. Maybe the center counter defense was not garbage after all. Alexander played carefully and powerfully. Then he went into deep thought over a position that looked quite dangerous for Deep Thought. It appeared to me that Alexander could unleash a very strong attack on Deep Thought’s king. To my surprise, Alexander played a move that seemed to release all the tension in the position. What was he doing? It looked like an even ending to me. On a closer look at the board, I began to get a sinking feeling. It wasn’t an even ending. Alexander had seen deeper than Deep Thought, and Deep Thought lost its first ever game. While Deep Thought was having its problems, interesting things were unfolding on ChipTest’s board. Before Vivek’s game with ChipTest, he was openly expressing his contempt of chess-playing computers, having had numerous pleasant memories at the expense of Hitech from earlier encounters. ChipTest forced Vivek to resign in twenty-two moves with an unexpected sacrifice. Hours after the game was over, Vivek was still showing other people ChipTest’s surprise combination.

在上一轮比赛中,ChipTest对阵人类棋手的不败纪录戛然而止,输给了一位强大的高手。尽管Vivek输给了ChipTest,但他仍然表达了对国际象棋计算机的蔑视,并在最后一轮得到了证明自己观点的机会。他与深思配对。现在受到 ChipTest 战术能力的考验,Vivek 与 Deep Thought 进行了一场更安静的比赛。达到了“孤立的皇后兵”位置。Deep Thought 在棋盘中央拥有更多控制权,但其孤立的皇后棋子也存在潜在弱点。Deep Thought 的后棋子在相邻的文件上没有任何友方棋子。如果维韦克能够在大部分棋子被交易的情况下推动游戏进入结局,他很可能会获胜,因为深思的皇后棋子将成为游戏终局的主要弱点。维韦克成功地交易了两枚棋子,似乎获得了有利的位置。然后深思想开始打出看似愚蠢的动作。维维克的脸上露出了笑容。他感觉到了杀意吗?深思依然漠不关心。维韦克继续他安静的动作,动作看起来天真无邪。然后,Deep Thought 将棋子推向 f 档,威胁要打开局面。它的分数跃升。维韦克愣住了,又看了一眼,然后离开房间呼吸新鲜空气。这是一个很好的举动。他回来坐在桌边,这次双手捂住耳朵。有人打开门,发出轻微的声音。维维克让那家伙安静下来。维维克思考了二十五分钟后,做出了行动。随后发生了一系列的棋子交换。十步之后,硝烟散尽,维韦克落下棋子。几次行动后,当他即将失去更多材料时,他辞职了。因此,宾夕法尼亚州的顶级球员在同一天两次输给宾夕法尼亚州最新的 USCF 球员。

In the last round, ChipTest’s undefeated record against human chess players came to a screeching halt, losing to a strong master. Despite Vivek’s loss to ChipTest, he continued to express his contempt of chess computers, and got a chance to prove his point in the last round. He was paired against Deep Thought. Now chastened by ChipTest’s tactical ability, Vivek went for a quieter game against Deep Thought. An “isolated queen pawn” position was reached. Deep Thought had more control in the center of the board, but also a potential weakness in its isolated queen pawn. Deep Thought’s queen pawn did not have any friendly pawns on the adjacent files. If Vivek could drive the game into an ending with most of the pieces traded off, he would very likely win, as Deep Thought’s queen pawn would become a major endgame weakness. Vivek managed to trade off two pieces, and seemed to have obtained a favorable position. Then Deep Thought started playing dumb looking moves. A smile formed on Vivek’s face. Did he sense a kill? Deep Thought remained unconcerned. Vivek continued his quiet maneuver with innocent looking moves. Then Deep Thought pushed its pawn on f file, threatening to open up the position. Its score jumped. Vivek froze up, took another look, and left the room for fresh air. It was a good move. He came back and sat at the table, this time with his hands over his ears. Someone opened the door and made a slight noise. Vivek hushed the guy. After thinking for twenty-five minutes, Vivek made his move. A series of piece exchanges ensued. After all the smoke cleared ten moves later, Vivek was down a pawn. A few moves later, when he was about to lose more material, he resigned. So the top player in Pennsylvania lost twice in the same day to the newest USCF players in Pennsylvania.

罗斯,我们的第一轮对手,在维韦克比赛后顺便过来了。他评论道:“大约 2200 秒。” 然后他补充说,如果他知道自己要面对什么,他就会接受抽签。

Ross, our first round opponent, dropped by after the Vivek game. He commented, “Some 2200s.” He then added that if he had known what he was dealing with, he would have just taken the draw.

Alexander Ivanov 赢得了该赛事,满分 6 分,获得 5 分。Deep Thought 拥有两位大师,并列第二名,满分 4.5 分。ChipTest 并列第五,满分 4 分。Hitech 的比赛很糟糕,得分 3.5满分 6 分。Deep Thought 在比赛中的表现评分为 2599,而 ChipTest 的评分为 2521。Hitech 的评分为 2312。

Alexander Ivanov won the event, scoring 5 points out of 6. Deep Thought tied for the second place with two masters, scoring 4.5 out of 6. ChipTest tied for fifth place, scoring 4 out of 6. Hitech had a bad tournament, scoring 3.5 out of 6. Deep Thought’s performance rating was 2599 for the tournament, while ChipTest’s was at 2521. Hitech performed at 2312.

如果ChipTest和Deep Thought可以被认为是一台机器,那么它的12场比赛成绩在2560左右。Fredkin中级奖的规则要求25场比赛成绩超过2500。我们有很好的机会,假设 Hitech 不会首先这样做。在 Fredkin Masters Open 之前,我不相信 Hitech 或 ChipTest 能够做到这一点。但现在看来ChipTest可以做到,谁能肯定Hitech也不能做到呢?我们有了新的紧迫感。Deep Thought 的射击能力是三台机器中最好的,但这将是一场比赛。我们拥有领先地位和更强大的机器。但不要忘记兔子在与乌龟赛跑时发生了什么。

If ChipTest and Deep Thought could be considered as a single machine, then its twelve-game performance would be around 2560. The rules for the Fredkin Intermediate Prize called for a twenty-five-game performance of over 2500. We had a very good shot, assuming that Hitech would not do it first. Before the Fredkin Masters Open, I did not believe that either Hitech or ChipTest could have done it. But it now looked like ChipTest could do it, and who could say for sure that Hitech could not do it too? We had a new sense of urgency. Deep Thought had the best shot of the three machines, but it would be a race. We had the lead and the stronger machines. But don’t forget what happened to the hare in the race against the tortoise.

谢谢你!谢谢你!

Thank you! Thank you!

不断将新机器引入国际象棋比赛的缺点之一是需要处理大量令人兴奋的事情等待发生的错误。ChipTest 在 1986 年 ACM 锦标赛上首次亮相,令人难忘。Deep Thought 0.01在Fredkin大师公开赛上的首次亮相当然也有其紧张的时刻,但没有什么明显的,至少从我们对手的角度来看是这样。Deep Thought 0.02 于 8 月 7 日至 19 日在波士顿举行的 1988 年美国公开赛上参加了第一场比赛。这次经历让人想起 ChipTest 的第一场比赛。

One of the drawbacks of constantly bringing new machines to chess competitions, is the need to deal with the numerous exciting bugs waiting to happen. ChipTest had a memorable first outing at the 1986 ACM Championship. The debut of Deep Thought 0.01 in the Fredkin Masters Open certainly had its tense moments, but there was nothing obvious, at least from our opponents’ viewpoints. Deep Thought 0.02 played its first tournament at the 1988 US Open held in Boston from August 7 to 19. It was an experience reminiscent of ChipTest’s first tournament.

Deep Thought 的绕线卡上有两个国际象棋处理器。在 Fredkin Masters Open 比赛中,只使用了其中一个处理器;没有时间让双处理器版本运行。Fredkin Masters Open 上使用的单处理器版本改编自 ChipTest,使用旧的搜索代码,但具有全新的评估功能。由于一些现已失传的古老原因,我们在美国公开赛前的两个月里没有费心去参加比赛。也许,我们想尽快获得双处理器版本。另一种可能性是我们想要一本像样的开篇书。

Deep Thought’s wire-wrapped card had two chess processors on it. At the Fredkin Masters Open, only one of the processors was in use; there was no time to get a dual-processor version working. The single-processor version that played at the Fredkin Masters Open was adapted from ChipTest, using the old search code but with the brand new evaluation function. For reasons now lost in antiquity, we did not bother to play in tournaments in the two months leading up to the US Open. Perhaps, we wanted to get the dual-processor version as soon as possible. The other possibility was that we wanted to get a decent opening book.

ChipTest 的开局书严重不足以与人类玩家对战。Murray Campbell 在 Fredkin Masters Open 之前获得了博士学位,并获得了计算机科学系的助理研究员职位。该职位相当于助理教授,但属于研究方向。柏林纳是他的半官方老板1。Fredkin Masters Open 结束后,Murray 被告知他不能花任何时间在 Deep Thought 或 ChipTest 上——毕竟,他受雇从事其他项目。默里在办公时间严格遵守指示,但在闲暇时间继续研究这两个项目。随着穆雷的时间成为稀缺资源,我们将新的开篇书移至待办事项列表的顶部。由于默里没有足够的时间来做这件事,我们向大学外部寻求帮助。国际大师拉里·考夫曼主动提出向我们提供他为雷克斯国际象棋项目准备的入门书,我们接受了他的提议。讽刺的是,Rex 是 1986 年 ACM 锦标赛上 ChipTest 的第一个受害者。

The ChipTest opening book was grossly inadequate for playing against human players. Murray Campbell received his PhD degree before the Fredkin Masters Open, and was offered the faculty position of Research Associate at the CS Department. The position was equivalent to Assistant Professor but in a research track. Berliner was his semi-official boss1. After the Fredkin Masters Open, Murray was told that he was not to spend any time on either Deep Thought or ChipTest—after all, he was employed to work on other projects. Murray followed the directive to the letter in office hours, but continued to work on the two programs in his leisure time. With Murray’s time becoming a scarce resource, we moved the new opening book to the top of the to-do list. Since Murray would not have enough time to work on it, we sought help from outside the university. International Master, Larry Kaufman, offered to provide us with the opening book that he had prepared for the Rex chess program, and we took him up on his offer. Ironically, Rex had been ChipTest’s first victim at the 1986 ACM Championship.

当我们推出新版本时,Hitech 参加了 1988 年 7 月 23 日至 24 日的宾夕法尼亚州锦标赛,赢得了冠军,最终评分超过 2400,成为第一家正式计算机高级硕士,诞生三年后2 . ChipTest和Deep Thought的评级都比Hitech高,但还没有玩过二十个评级的游戏,这是正式高级大师的最低要求。虽然没有与高级大师头衔相关的奖项,但 Deep Thought 团队成员仍存在一些疑虑,认为我们应该向锦标赛发送 Deep Thought 的单处理器版本。无论如何,Hitech这只乌龟正在迎头赶上,我们不能再等了。

While we were bringing up the new version, Hitech played in the 1988 Pennsylvania State Championship from July 23–24, won the title, finally getting its rating over 2400, and became the first official computer senior master, three years after its birth2. Both ChipTest and Deep Thought were higher rated than Hitech, but had not yet played twenty rated games, the minimum needed to be official senior masters. While there was no prize associated with the senior master title, there were some misgivings among the Deep Thought team members, that we should have sent out the single-processor version of Deep Thought to tournaments. Anyhow, Hitech the tortoise was catching up, and we could not wait any longer.

美国公开赛是美国国际象棋联合会的展示赛事,是在波士顿举行的十二轮比赛。托马斯·阿南塔拉曼 (Thomas Anantharaman) 有一位朋友可以和他一起留在波士顿,因此他自愿开车前往波士顿,并担任接线员。托马斯的费用将从我们之前在 ACM 锦标赛和弗雷德金大师公开赛中获得的奖金来支付。由于常规国际象棋比赛不允许计算机获得奖金,因此我们不会再获得任何奖金,并且必须注意我们花了多少钱。我们将第一次使用双处理器版本。

The US Open, the showcase event for the US Chess Federation, was a twelve-round tournament held in Boston. Thomas Anantharaman had a friend with whom he could stay in Boston, so he volunteered to drive to Boston, and be the operator. Thomas’ expenses would be paid from our earlier winnings in the ACM Championship and the Fredkin Masters Open. As computers were not allowed to get the prize money in regular chess tournaments, we would not be getting any more winnings, and had to watch how much we spent. We would be playing with the dual-processor version for the first time.

美国公开赛的前三场比赛都是使用双处理器代码进行的,结果是灾难性的。第一个游戏的绘制是由于硬件和软件代码之间的接口错误造成的,尽管我们当时并不确定原因。Deep Thought 0.02在对手提前认输的情况下赢得了第二局,但如果他知道第三轮会发生什么,他可能会继续比赛。在第三轮中,我们遇到了一位专家级的对手,当他的位置明显丢失时,他干脆拒绝辞职。Deep Thought 0.02 被车领先,但对手继续比赛。然后奇怪的事情开始发生。Deep Thought 0.02开始玩无意义的动作。对手下棋威胁要在下一步将其将死,而深思 0.02 没有费心去防御这一威胁。对手刚刚做出交配动作,便惊呼道:“谢谢!谢谢你!” 当他逃跑时。

The first three games at the US Open were played with the dual-processor code, and the results were disastrous. The first game was drawn as a result of a bug at the interface between the hardware and the software code, although we were not certain of the cause at the time. Deep Thought 0.02 managed to win the second game when the opponent resigned early, but had he known what was to happen in the third round, he probably would have played on. In the third round, we got an expert-level opponent who simply refused to resign when his position was clearly lost. Deep Thought 0.02 was up by a rook, but the opponent played on. Then strange things started to happen. Deep Thought 0.02 started playing nonsensical moves. The opponent made a move threatening to checkmate it on the next move, and Deep Thought 0.02 did not bother to defend against the threat. The opponent just made the mating move, and then exclaimed “Thank you! Thank you!” as he ran off.

事实证明,托马斯的新双处理器代码以与单处理器代码不同的方式访问硬件。这有导致硬件损坏的副作用在某些条件下反转将死值的符号的微程序。Deep Thought 并没有避免被将死,而是享受被将死!

It turned out that Thomas’ new dual-processor code was accessing the hardware in a different way from the single-processor code. This had the side effect of causing the hardware microprogram to reverse the sign of the value for checkmates under certain conditions. Instead of avoiding being checkmated, Deep Thought was enjoying being checkmated!

由于托马斯在波士顿,调试双处理器代码很困难,第三轮后我们恢复到更可靠的单处理器程序。接下来的五场比赛,深思0.01只让一场平局,其余四场均获胜。

With Thomas in Boston, it was difficult to debug the dual-processor code, and we reverted to the more reliable single-processor program after round three. In the next five games, Deep Thought 0.01 allowed one draw, winning the other four.

第九轮迎来了第一个主要对手,国际大师伊戈尔·伊万诺夫。(几年后,我拿起一本关于世界冠军加里·卡斯帕罗夫(Garry Kasparov)早期职业生涯的书,并惊讶地读到伊戈尔和加里并列第一的比赛。但我有点超前了。)行动要早得多,甚至在 ChipTest 出现之前。这是他和 Hitech 之间的一场锦标赛,比赛地点是匹兹堡大学校园,距离我的办公室步行十分钟。在那盘棋中,伊戈尔毫不费力地布置棋子进攻,屠杀了海泰克。Hitech 小组的戈登·戈奇 (Gordon Goetsch) 指出,这款游戏是计算机尚未发展到何种程度的一个例子。ChipTest 构建完成后,我测试了 Igor 与 Hitech 比赛中的位置,并对他的表现印象深刻。ChipTest可能比Hitech早几招就看到了Hitech的灭亡,但也不知道他在做什么。伊戈尔有宗师实力,但却没有宗师称号。他在西方举行的一场锦标赛期间从苏联叛逃,这可能使他很难获得冠军头衔。他是加拿大并列冠军,美国公开赛成绩为 2641 分,是公开赛上的顶尖选手之一。

Round nine brought the first major opponent, International Master Igor Ivanov. (Years later, I picked up a book on the early career of World Champion Garry Kasparov, and was surprised to read about a tournament where Igor and Garry tied for first place. But I am getting ahead of myself.) I had seen Igor in action much earlier, even before ChipTest’s existence. It was a tournament game between him and Hitech, played on the University of Pittsburgh campus, a ten-minute walk from my office. In that game, Igor effortlessly arranged his pieces for an attack and slaughtered Hitech. Gordon Goetsch of the Hitech group pointed to this game as an example of how far computers had yet to go. After ChipTest was built, I tested out positions from the Igor vs. Hitech game and was quite impressed with his play. ChipTest probably saw Hitech’s demise a few moves earlier than Hitech did, but also had no idea what he was doing. Igor was of Grandmaster strength, but he did not have a Grandmaster title. He had defected from the USSR during a tournament held in the West, and this might have made it difficult for him to obtain the title. He was the Canadian co-champion, and at 2641 on the US scale was one of the top players at the Open.

Deep Thought 会如何对抗伊戈尔·伊万诺夫?在匹兹堡,我们显然很紧张。在波士顿,托马斯可能对他的历史一无所知,但仅他的评分就足以给人留下深刻的印象。伊戈尔有黑色。后来,托马斯提到伊戈尔在游戏的早期阶段正在看一本杂志;他对深思毫无尊重。弗雷德金大师公开赛的结果在国际象棋界仍然鲜为人知。

How would Deep Thought do against Igor Ivanov? In Pittsburgh we were obviously nervous. In Boston, Thomas was probably ignorant of his history but his rating alone would be sufficient to impress. Igor had Black. Later, Thomas mentioned that Igor was reading a magazine during the early part of the game; he had no respect whatsoever for Deep Thought. The results from the Fredkin Masters Open were still largely unknown in the chess world.

对阵伊戈尔的比赛变成了我见过的最奇怪的事情之一。深思喜欢推动它的棋子。用国际象棋术语来说,它喜欢增加空间。优秀的国际象棋棋手在早期阶段就被教导,将棋子推得太远可能会造成弱点。“超现代”的国际象棋学派是故意挑衅对手推动棋子,希望棋子变得过度延伸并且难以维持。如果这就是伊戈尔的意图,那么他就取得了辉煌的成功,至少在令人兴奋的部分是这样。到第十九步时,深思的所有棋子都前进了至少一格;事实上,除了边缘的两个棋子之外,其他棋子都至少前进了两格。由于 Deep Thought 放任不管,我很担心,但它真的很喜欢它的位置。发生了什么事?据托马斯说,此时伊戈尔已经放下杂志进行了几次动作;他看到了一些他不喜欢的东西。深思的下一步棋将他所有的棋子都推回了前两列,他几乎无法移动。它的棋子并没有过度扩张——他们占据了主导地位。两步棋后,伊戈尔的处境恶化并变得至关重要。第二十九步,他在物质即将遭受重大损失时辞职了。

The game against Igor turned into one of the strangest affairs that I have ever seen. Deep Thought liked to push its pawns. In chess terms, it liked to increase space. Good chess players are taught at an early stage that pushing pawns too far could create weaknesses. The “hypermodern” chess school of thought is to deliberately provoke the opponent to push their pawns, hoping that the pawns become over-extended and difficult to maintain. If that was Igor’s intent, he succeeded brilliantly, at least for the provoking part. By move nineteen, all of Deep Thought’s pawns had advanced at least one square; in fact, with the exception of the two pawns on the edges, all of them had advanced by at least two squares. With Deep Thought pushing pawns with abandon, I was quite concerned, but it really liked its position. What was going on? According to Thomas, by this time, Igor had put down his magazine for several moves; he had seen something that he did not like. Deep Thought’s next move pushed all of his pieces back to the first two ranks, and he could hardly move. Its pawns were not over-extended—they were dominating the position. Two moves later, Igor’s position deteriorated and became critical. On move twenty-nine, he resigned just as he was about to suffer heavy material loss.

1987年美国公开赛冠军、特级大师列夫·阿尔伯特(Lev Alburt)的评分为2652,略高于伊戈尔(Igor),他在第十轮对阵“Deep Thought”。伊戈尔(Igor)输给“深思”(Deep Thought)是一个很大的遗憾,列夫(Lev)被充分警告他将面临什么。列夫打得很谨慎,进入残局时形势一片大好。第十九步,深思的位置变得至关重要,它将棋子推到了国王前面两格。列夫对这一举动感到惊讶,这可能是为深思保住位置的唯一举动。虽然 Deep Thought 的处境看起来很糟糕,但它并没有失败。然而,狡猾的列夫却找到了取得进步的方法。比赛随后变成了车结局,双方都只有一车和五兵。列夫仍然占据优势,但众所周知,新车结局很难获胜。也许我们可以打平。深思确实在第五十二步手上有重复抽牌,如果它已经下过抽牌就好了。由于未知的原因,深思找到了该招式,但随后决定考虑另一种绘图招式。然而,这一新举措彻底失败了。Deep Thought发现新的绘图动作不起作用,而是控制时间分配的软件不够智能,无法让它切换回真正的绘图动作。我们所有人都痛苦地看着“深思”花了二十分钟试图找到它已经知道的绘图动作。它没有重新发现这一步棋,我们在第 61 步棋中放弃了计算机。

The 1987 US Open Champion, Grandmaster Lev Alburt, rated slightly higher at 2652 than Igor, was paired against Deep Thought in the tenth round. Igor’s loss to Deep Thought had been a big upset, and Lev was adequately warned about what he would be facing. Lev played cautiously and entered the endgame with a promising position. On move nineteen, Deep Thought’s position became critical, and it pushed a pawn in front of its king by two squares. Lev was surprised by the move which was probably the only move that held the position for Deep Thought. While Deep Thought’s position looked ugly, it was not losing. The wily Lev, however, found a way to make progress. The game then became a rook ending where both sides had only one rook and five pawns each. Lev still held the edge, but rook endings are notoriously difficult to win. Perhaps we could get a draw. Deep Thought indeed had a repetition draw in hand by move fifty-two, if only it had played the drawing move. For reasons unknown, Deep Thought found the move, but then decided to look at an alternative drawing move. The new move, however, lost outright. Deep Thought found that the new drawing move did not work, but the software controlling the time allocation was not smart enough to allow it to switch back to the real drawing move. All of us watched in agony while Deep Thought spent twenty minutes trying to find the drawing move that it already knew about. It did not rediscover the move, and we resigned for the computer on move sixty-one.

深思在下一场比赛中战胜了一位高手,然后在最后一轮比赛中与一位高级高手打平。它是临时评级从 2599 下降到 2495,主要是由于前三轮的 bug 表现。如果我们把这三场比赛拿出来,只计算Deep Thought的比赛场次为0.01,那么它的十五场比赛(其中六场来自弗雷德金大师赛,九场来自美网最后几轮)的表现约为2592。如果Deep Thought 0.01与Lev打平,那么其十五场比赛的成绩将在2634左右。

Deep Thought won the next game against a master player, and then drew the final round game against a senior master. Its provisional rating dropped from 2599 to 2495, mostly as a result of its buggy performance in the first three rounds. If we take out the three games in question, and only count the games played by Deep Thought 0.01, then the performance for its fifteen games (six from the Fredkin Masters Open, and nine from the final rounds of the US Open) would be around 2592. Had Deep Thought 0.01 drawn the game against Lev, its fifteen-game performance would be around 2634.

美国公开赛是一次发人深省的经历。首先,我们有一些错误需要修复。其次,弗莱德金中级奖显然是触手可及的,但问题是海泰是否能打败我们。

The US Open was a sobering experience. First, we had bugs to fix. Second, the Fredkin Intermediate Prize was obviously within reach, but the question was whether Hitech would beat us to it.

Hitech 在比赛中并没有做太多的事情,所以我们有一些时间重新组合。我们在接下来的两个月里修复了 Deep Thought 0.02 中的错误。在这个过程中,我们还发现了另一个重大bug。Deep Thought 没有为微程序控制的搜索部分生成en passant 3步。让我们困惑了一段时间的是,尽管 Deep Thought 0.01 的国际象棋知识比 ChipTest 多得多,但在面对面的比赛中,Deep Thought 0.01 的表现并不比 ChipTest 好。路过bug 可能是罪魁祸首。

Hitech was not doing much tournament-wise, so we had some time to regroup. We spent the next two months ironing out the bugs in Deep Thought 0.02. During the process, we found another major bug. Deep Thought was not generating en passant3 moves for the part of the search controlled by the microprogram. It had been puzzling us for a while that Deep Thought 0.01 was not playing better than ChipTest in head-to-head competition, despite having far more chess knowledge than ChipTest. The en passant bug might be the main culprit.

1988 年 10 月,Deep Thought 0.02 又参加了两场比赛:10 月 8 日至 10 日在加利福尼亚州洛杉矶举行的美国公开赛,以及 10 月 28 日至 30 日在新泽西州萨默塞特举行的美国级锦标赛。斯图尔特·克拉克拉夫特(Stuart Cracraft)和吉姆·吉洛格利(Jim Gillogly)在美国公开赛上慷慨地担任我们的操作员。安德烈亚斯前往新泽西州,在那里 Deep Thought 又击败了两位国际大师。美国级锦标赛结束后,Deep Thought 已经打了 29 场额定比赛。二十九场比赛(包括三场受交配错误影响的比赛)的整体表现为 2509。最好的二十五场比赛表现(同样包括三场糟糕的比赛)为 2519。换句话说,我们本可以在美国组冠军赛后立即获得 Fredkin 中级奖。然而,我们并不知道 Deep Thought 已经实现了这一里程碑;我们只知道根据挂图收视率,它已经很接近了,而挂图已经过时了几个月。不知道我们已经赢得了比赛结果是一件好事。接下来的一个月,“深度思考”即将震惊国际象棋界。

Deep Thought 0.02 played in two more tournaments in October 1988, the American Open from October 8 to 10 in Los Angeles, California, and the US Class Championship from October 28 to 30 in Somerset, New Jersey. Stuart Cracraft and Jim Gillogly graciously served as the operators for us in the American Open. Andreas traveled to New Jersey where Deep Thought defeated two more International Masters. After the US Class Championship, Deep Thought had played twenty-nine rated games. The overall performance for the twenty-nine games, including the three games marred by the mating bug, was 2509. The best twenty-five-game performance, again including the three bad games, was 2519. In other words, we could have claimed the Fredkin Intermediate Prize right after the US Class Championship. However, we had no idea that Deep Thought had already achieved the milestone; we only knew that it was close, based on the wall-chart ratings, which were a few months out of date. Not knowing that we had already won the race turned out to be a good thing. The following month Deep Thought was about to shock the chess world.


鉴于 ChipTest 在美国公开赛前的正面交锋中略领先 Deep Thought 0.01,如果我们让它参加人类锦标赛的话,它完全有可能在 1987 年获得 Fredkin 中级奖。但那样的话,“深度思考”的发展就会受到影响。

Given that ChipTest was slightly ahead of Deep Thought 0.01 up to the US Open in head-to-head competitions, it was entirely possible that it could have claimed the Fredkin Intermediate Prize in 1987, if only we had let it play in human tournaments. But then, the development of Deep Thought would have suffered.

领先世界冠军

Ahead of a World Champion

1988 年 1 月,在巴黎举行的新闻发布会上,国际象棋世界冠军加里·卡斯帕罗夫被问及计算机能否在 2000 年之前的锦标赛中击败特级大师。我很乐意提供我的建议。” 加里不知道同一年会发生什么。他也没有想到,九年后的1997年,也就是三年前的2000年,他自己将面临什么样的困难。

In January 1988, at a press conference in Paris, World Chess Champion Garry Kasparov was asked whether a computer could defeat a Grandmaster in tournament play before the year 2000. He replied, “No way, and if any Grandmaster has difficulties playing computers, I would be happy to provide my advice.” Garry had no idea what was about to happen in the very same year. Nor did he have any inkling as to what kind of difficulties he himself would be facing nine years later in 1997, three years before the year 2000.

当加里做出这一预测时,这似乎是一个安全的预测。十多年前,计算机确实在闪电战(五分钟)游戏中击败了特级大师,特级大师迈克尔·斯蒂恩和特级大师罗伯特·许布纳成为受害者。但在锦标赛条件下,计算机从未与特级大师打平过一场比赛。虽然 ChipTest 已经发挥了全部实力,但它在计算机国际象棋界之外并不为人所知。事实上,就连ChipTest团队也并不知道它有多强,因为该程序直到Fredkin大师公开赛才获得评级。Cray Blitz 曾是世界电脑象棋冠军,但国际大师 David Levy 在 1984 年的一场比赛中羞辱了它,赢得了全部四场比赛。在美国,评分最高的国际象棋计算机 Hitech 多年来仅徘徊在 2350 到 2400 之间。在同一等级中,特级大师的评分通常高于 2550。毫无疑问,加里的预测是正确的。

At the time Garry made the prediction, it seemed to be a safe one. Computers had indeed beaten Grandmasters in blitz (five-minute) games more than a decade earlier, with Grandmaster Michael Stean and Grandmaster Robert Huebner as the victims. But under tournament conditions, computers had never drawn even a single game against a Grandmaster. While ChipTest was already in its full strength, it was unknown outside the computer chess community. As a matter of fact, not even the team really knew how strong ChipTest was, since the program did not have a rating until the Fredkin Masters Open. Cray Blitz was the World Computer Chess Champion, but International Master David Levy humiliated it in a 1984 match, winning all four games. The highest rated chess computer, Hitech, for years only hovered between 2350 and 2400 on the US scale. Grandmasters were typically rated above 2550 on the same scale. Surely, Garry’s prediction had the ring of truth to it.

1988年,名人堂国际象棋节锦标赛于11月4日至6日在俄亥俄州坎顿举行。Deep Thought 进行了比赛迈克·布朗担任操作员。Deep Thought 的比赛表现出色,赢得了前四轮比赛,其中还战胜了另一位国际大师。第五轮,Deep Thought 在其短暂的职业生涯中第二次对阵伊戈尔·伊万诺夫。伊戈尔有怀特,这次没有读任何杂志!比赛以平局告终。Deep Thought 首次赢得锦标赛胜利,与 Igor 并列第一。

In 1988, the Hall of Fame Chess Festival tournament was held in Canton, Ohio, from November 4 to 6. Deep Thought played with Mike Browne as the operator. Deep Thought had a good tournament, winning the first four rounds, including one over another International Master. In the fifth round, Deep Thought played Igor Ivanov for the second time in its short career. Igor had White and this time did not do any magazine reading! The game ended in a fighting draw. Deep Thought had its first tournament win, as it tied for first with Igor.

下周末,Deep Thought 在佛罗里达州奥兰多参加了首届 ACM 锦标赛。Deep Thought 在抢七中击败了首轮战平的 Fidelity Challenger,赢得了冠军。在本次比赛中,Deep Thought 首次与 Hitech 公开比赛并获胜。(Deep Thought 在计算机国际象棋比赛中与 Hitech 进行了 3 次比赛,其中一场输了,另外两场获胜。两个程序之间 3 比 1 的胜率与程序之间大约 200 个评分点的划分一致。)它所参加的所有计算机象棋锦标赛的明显获胜者或共同获胜者。但运气在结果中发挥了重要作用。考虑到这些锦标赛的瑞士锦标赛形式,Deep Thought 尽管是最受欢迎的选手,但赢得每项单项赛事的机会只有大约 50%。

The next weekend, Deep Thought played in its first ACM Championship in Orlando, Florida. Deep Thought won the title on tiebreak over Fidelity Challenger, which it drew in the first round. In this tournament Deep Thought played Hitech in public for the first time and won. (Deep Thought played Hitech three more times in computer chess events, lost one of them and won the other two. The winning ratio of 3 to 1 between the two programs is consistent with the roughly 200 rating points separating the programs.) Deep Thought was the clear winner or the co-winner of all the Computer Chess Championships that it played in. But luck played a big part in the results. Given the Swiss tournament format of these Championships, Deep Thought, despite being the heavy favorite, had only about fifty-percent chance of winning each individual event.

11 月 24 日至 27 日,Software Toolworks 锦标赛在加利福尼亚州长滩举行。总奖金达13万美元,是1988年美国三大国际象棋锦标赛之一。除了几位实力强劲的美国特级大师和国际大师之外,还邀请了三位传奇人物:前世界冠军米哈伊尔·塔尔、前世界冠军候选人“伟大”戴恩·本特·拉森 (Dane Bent Larsen) 和前世界冠军候选人塞缪尔·雷舍夫斯基 (Samuel Reshevsky)。米哈伊尔和本特仍然很活跃,并且在世界收视率排行榜上名列前茅,分别位列世界第十六和第四十二。

Between November 24 and 27, the Software Toolworks Championship was held in Long Beach, California. With a total prize fund of $130,000, it was one of the top three US chess tournaments in 1988. Besides several strong US Grandmasters and International Masters, there were three invited legends: the former World Champion Mikhail Tal, the former World Championship Candidate “Great Dane” Bent Larsen, and the former World Championship Candidate Samuel Reshevsky. Both Mikhail and Bent were still active and ranked high on the world-rating list, sixteenth and forty-second in the world, respectively.

Stuart Cracraft 和 Jim Gillogly 再次为远程操作员提供了宝贵的帮助。Jim 碰巧认识 Software Toolworks(现在称为 Mindscape)的创始人之一,他在 Toolworks 人员的帮助下通过电话线解决了问题。时任卡内基梅隆大学计算机科学系高级教授 Raj Reddy 教授4主动提出为 Deep Thought 支付比赛费用。(过了很多年我才知道 Raj 最初对 Fredkin 奖的设立也发挥了重要作用。)到比赛时,我们知道 Deep Thought 可能已经赢得了 Fredkin 中级奖,但还有一些关于如何计算“绩效”的模糊性。如果使用数学上不正确的公式,“深度思考”就会失败。在这种情况下,我们还需要两场精彩的比赛。

Again Stuart Cracraft, and Jim Gillogly offered their invaluable help to be the remote operators. Jim, who happened to know one of the founders of Software Toolworks (known as Mindscape these days), smoothed the problems with the phone line with the help of Toolworks’ personnel. Professor Raj Reddy, then a senior professor in the Computer Science Department at Carnegie Mellon 4, offered to foot Deep Thought’s bill for the tournament. (It would be many years before I knew that Raj was also instrumental in setting up the Fredkin Prize in the first place.) By the time of the tournament, we knew that Deep Thought might have already won the Fredkin Intermediate Prize but there was some ambiguity on how the “performance” should be calculated. Deep Thought would have fallen short if the mathematically incorrect formula were to be used. We would have needed two more good games in that case.

虽然 Deep Thought 已经参加了半年的锦标赛,但国际象棋界大多不知道它的存在。当比赛总监听到 Deep Thought 的评级时,他再三考虑,并打电话给美国国际象棋联合会以确保其正确性。我的一位老熟人桑乔伊·马哈詹 (Sanjoy Mahajan) 在参加锦标赛时偶然遇到了 IM Patrick Wolff,他后来成为了特级大师和美国冠军。帕特里克询问 Deep Thought 在比赛开始时的评分是多少。桑乔伊说是 2510 年,帕特里克对此说:“不可能有那么强。” 随着比赛的进行,帕特里克将大大改变他的态度。

While Deep Thought had played in tournaments for half a year, the chess world was mostly unaware of its existence. The tournament director did a double take when he heard Deep Thought’s rating and phoned the US Chess Federation to make sure that it was correct. An old acquaintance of mine, Sanjoy Mahajan, was playing in the tournament and chanced on IM Patrick Wolff, who later became a Grandmaster and US Champion. Patrick asked what Deep Thought’s rating was at the beginning of the tournament. Sanjoy said it was 2510, to which Patrick said, “No way is it that strong.” Patrick was to change his tune considerably as the tournament progressed.

Deep Thought 没有遇到太多麻烦就赢得了前两场比赛。就弗​​莱德金中级奖而言,我们已经如鱼得水了。斯图尔特是第二场比赛的操作员。国际象棋桌上的正常礼仪是不宣布自己即将将死对手。上世纪的老大师们以宣布将死而闻名,但今天这被认为是不好的形式。比赛结束时,深思计算出自己可以将敌19步,即深思19步加上对手回击18步,总共37层。斯图尔特非常兴奋,他告诉了对手。对方立即辞职。即使没有立即交配的可能性,这个位置也完全丧失了。

Deep Thought won the first two games without much trouble. As far as the Fredkin Intermediate Prize was concerned, we were home and dry. Stuart was the operator in the second game. The normal etiquette at the chess table is that one does not announce that one is about to checkmate the opponent. The old masters from the last century were known to announce checkmates, but it is considered bad form today. At the end of the game, Deep Thought calculated that it could checkmate its opponent in 19 moves, that is, 19 moves by Deep Thought plus 18 replies by the opponent, for a total of 37 plies. Stuart was so excited that he told the opponent. The opponent resigned promptly. The position was quite lost even without the immediate mating possibility.

第二天,我们从加州得到消息,我们第三轮的对手不是别人,正是“大丹犬”本特·拉森本人。本特是战后最成功的大师之一。他不仅是一次世界锦标赛候选人,而且在四个不同的世界锦标赛周期中都是世界锦标赛的候选人。在1970年举行的“苏联对世界其他国家”团体赛中,他领先于鲍比·费舍尔为世界其他国家打了第一板。在巅峰时期,他曾一度排名世界第二。在赢得 Fredkin 中级奖的兴奋之后,是时候进行现实检验了。本特会拥有白棋。我抱着最好的希望,也做好最坏的打算。我也有点羡慕斯图尔特和吉姆。人们每天都无法亲眼目睹一位传奇人物,更不用说面对面了。

The next day, we got word from California that our third round opponent was none other than the “Great Dane”, Bent Larsen himself. Bent was one of the most successful Grandmasters in the postwar era. He has been a World Championship Candidate, not just once, but in four different World Championship cycles. In “the USSR vs. the Rest of the World” team match held in 1970, he played the first board for the Rest of the World, ahead of Bobby Fischer. At his peak, he had been as high as number 2 in the world. After the thrill of winning the Fredkin Intermediate Prize, it was time for a reality check. Bent would have White. I was hoping for the best and bracing myself for the worst. I was also a little bit envious of Stuart and Jim. One does not get to see, let alone face, a legend every day.

本特并不是一个因循守旧的人,即使面对电脑,他也能下不寻常的棋。这场比赛前几年,我读过一本传记,其中提到他偏爱推动车棋子,即棋盘外侧边缘的两个棋子。Deep Thought 在其第一场比赛中也表现出了类似的趋势,但我很确定它的推理与 Bent 的推理完全不同。与本特的形式一致,在这场游戏中,他允许深思拥有一个先进的中心棋子,并且他确实将两个车棋子推到了任何一个中心棋子之前。即使过了这么多年,也很难判断这个位置。一位真人解说员断言本特在比赛开始阶段后具有优势。Deep Thought 对自己的地位非常满意。它有一个先进的棋子,限制了本特的位置。反过来,他对先进的棋子也产生了相当大的压力。对抗人类玩家时,对于 Bent 来说这可能确实是一个不错的位置,因为人类玩家可能在处理棋子上的压力方面存在问题。计算机会冷静地计算所有的变化,确保典当不会丢失,或者至少在没有类似补偿的情况下不会丢失。随着比赛的进行,Deep Thought 的评价也越来越有利于它。我开始思考一些不可思议的事情。Deep Thought真的有更好的地位吗?

Bent was not known as a conformist, and he played unusual chess even when facing a computer. A few years before this game, I had read a biography that mentioned his predilection for pushing the rook pawns, the two pawns on the outside edges of the chessboard. Deep Thought had showed a similar tendency in its first tournament, but I am quite sure that it had entirely different reasoning from Bent’s. True to Bent’s form, in this game, he allowed Deep Thought to have an advanced center pawn, and he did push both of his rook pawns before either of his center pawns. It is difficult to judge the position even after all these years. A human commentator asserted that Bent had an edge after the opening phase of the game. Deep Thought was quite happy about its position. It had an advanced pawn that cramped Bent’s position. In return, he had considerable pressure on the advanced pawn. Against a human player, it might indeed be a decent position for Bent, as human players might have problems handling the pressure on the pawn. A computer would just calmly calculate all the variations, making sure that the pawn is not lost or at least not without exacting comparable compensation. As the game progressed, Deep Thought’s evaluation climbed up in its favor. I was beginning to think the unthinkable. Could Deep Thought indeed have the better position?

最近的分析显示,使用更强大的国际象棋计算机(Deep Blue Jr),Deep Thought 可以下一个组合,并可能在第 17 步5上彻底赢得比赛。Deep Thought没有玩这个组合,但位置可能仍然有利。人类分析师普遍认为本特在比赛的早期占据了更好的位置。我多年来的观察是,只有真正优秀的分析师才能在分析人类与计算机游戏时保持客观。

Recent analysis, with a more powerful chess computer (Deep Blue Jr), showed that Deep Thought could have played a combination and possibly won the game outright on move seventeen5. Deep Thought did not play the combination, but the position was probably still favorable. The human analysts were universally of the opinion that Bent had the better position throughout the early part of the game. My observation over the years is that only really good analysts can stay objective when analyzing human vs. computer games.

慢慢地,本特扩大了他作品的范围。这个位置对两名球员来说都非常危险,对他来说可能更是如此。然后在第二十六步,本特犯了一个小错误。他动了他的国王,正好挡住了他主教撤退的广场。深思的下一步行动威胁要困住主教。本特本可以再次移动他的国王,为主教创造一个逃生广场,尽管会损失时间。几周后接受 IM David Levy 采访时,Bent 提到国王棋步是“一个非常糟糕的棋步”。事实上,糟糕的一步就是下一步。他不愿承认棋盘上的错误,于是将深思的棋子向前推了一把,猛烈攻击了深思的国王。这个位置对他来说已经接近失败,而在他下一步行动之后,深思算出了一场强行获胜。十五步棋后,本特辞职并成为第一位在锦标赛中在常规时间控制下输给电脑的特级大师。在匹兹堡,我们难以置信地看着对方。我们情绪高涨。但我们会在几个小时内回到现实。

Slowly, Bent improved the scope of his pieces. The position was very dangerous for both players, probably more so for him. Then on move twenty-six, Bent made a slight error. He moved his king and it happened to block the retreating square for his bishop. Deep Thought threatened to trap the bishop with its next move. Bent could have just moved his king again to create an escape square for the bishop, although with a loss of time. When interviewed by IM David Levy a few weeks later, Bent mentioned the king move as “one very bad move”. In reality, the bad move was the next one. Not willing to concede the mistake over the board, he lashed out at Deep Thought’s king by pushing one of his pawns forward. The position was close to losing for him, and after his next move, Deep Thought calculated a forced win. Fifteen moves later, Bent resigned and became the first Grandmaster to lose to a computer under regular time control in a tournament. In Pittsburgh, we looked at each other in disbelief. We were on an emotional high. But we would come down to earth within a few hours.

第四轮,Deep Thought对阵六届美国冠军特级大师Walter Browne。沃尔特以前曾在公共场合与电脑对战。1978年,他输给了西北国际象棋程序Chess 4.6,但同时他还与另外43个对手对弈!然而,这一次,他是全神贯注地与深思对战。国际象棋中有一句老话:“有时你给他们教训,有时他们给你教训。” 在这场比赛中,沃尔特给了深思一个很好的教训。当时的 Deep Thought 并不理解主教控制“开放对角线”的价值或“永久钉子”的价值6,沃尔特巧妙地利用了弱点,相当轻松地赢得了比赛。深思进行了相当大的斗争,但这还不够。

In the fourth round, Deep Thought played against six-time US Champion Grandmaster Walter Browne. Walter had played against computers in public before. In 1978, he lost a game against the Northwestern chess program, Chess 4.6, but he was also playing against forty-three other opponents at the same time! This time, however, he was playing against Deep Thought with undivided attention. There is an old saying in chess, “Sometimes you give them lessons, and sometimes they give you lessons.” In this game Walter gave Deep Thought a good lesson. Deep Thought at the time did not understand the value of a bishop controlling an “open diagonal” or the value of a “permanent pin”6, and Walter exploited the weaknesses beautifully and won the game fairly easily. Deep Thought put up quite a struggle, but it was not enough.

Deep Thought在第五轮对阵文斯·麦坎布里奇(Vince McCambridge)的比赛中,有史以来第二次以7比7逼平IM ,然后在接下来的比赛中战胜了一位强大的高手。第七轮,它的对手是新晋 IM Alex Fishbein。几年前,在 ACM 锦标赛上,IM Danny Kopec 带来了一篇他写的关于结局车和主教 vs. 车的文章。根据国际象棋残局书籍,这是一个典型的绘制结局。然而,丹尼·科佩克的文章提到,即使是大师也经常在这个结局中犯了输掉比赛的错误。由于 ChipTest 可以毫无问题地从 Danny 的获胜位置样本中找到交配序列,因此我们有意识地决定让 ChipTest 以及后来的 Deep Thought 参与结局。在与 Alex Fishbein 的比赛之前,Deep Thought 已经两次从这个结局的强方对阵专家级玩家的比赛中获胜,这个决定似乎是正确的。当“深思”让亚历克斯从明显失败的位置转变为车和主教对车结局的弱侧时,我对这个决定感到非常遗憾。理论上来说,棋盘上的位置是完美的平局。

Deep Thought drew with an IM for the second time ever7 in the fifth round game against Vince McCambridge, and then won the next game against a strong master. In the seventh round, it was paired against new IM Alex Fishbein. A couple of years ago, at the ACM Championship, IM Danny Kopec brought in an article he wrote about the ending rook and bishop vs. rook. According to chess endgame books, this is a typically drawn ending. Danny Kopec’s article, however, mentioned that even Grandmasters frequently made game-losing mistakes in this ending. Since ChipTest had no problem finding the mating sequence from Danny’s sample winning positions, we made a conscious decision to allow ChipTest, and later, Deep Thought, to play into the ending. Prior to the game against Alex Fishbein, Deep Thought had won twice from the stronger side of this ending against players of expert level and this decision seemed to be sound. I greatly regretted the decision when Deep Thought allowed Alex to convert from a clearly lost position into the weak side of rook and bishop vs. rook ending. Theoretically, the position on the chessboard was a draw with perfect play.

Sanjoy Mahajan 与高级大师 Adam Leif 和 IM Patrick Wolff 进行了交谈,他们都认为 Alex 应该知道如何演奏它。“他就是那种知道这一点的球员,”帕特里克说。有一段时间,亚历克斯保持着平局。由于最后一轮比赛即将到来,赛事总监将比赛推迟到最后一轮比赛之后进行。赛事总监之间就最后一轮如何搭配 Deep Thought 进行了一场有趣的辩论。如果这场比赛是 Deep Thought 获胜的话,我们上一轮的对手将是英国特级大师安东尼·迈尔斯。相反,这场比赛是平局,我们最终对阵 IM Jeremy Silman,他恰好是 Stuart Cracraft 的国际象棋老师,Stuart 面对自己的老师很紧张。当然,深思并不知道杰里米·西尔曼是谁。杰里米成功地将比赛带入了平局,但随后以五十七步落败,结局艰难。

Sanjoy Mahajan talked with senior master Adam Leif and IM Patrick Wolff, who both thought Alex should know how to play it. “He’s just the kind of player who’d know it,” said Patrick. For a while, Alex was maintaining the draw. Since the final round was coming up, the tournament directors adjourned the game to be played after the last round game. There was an interesting debate among the tournament directors about how Deep Thought should be paired in the last round. If the game had been paired as a win for Deep Thought, our last round opponent would have been British Grandmaster Anthony Miles. Instead, the game was paired as a draw, and we ended up playing IM Jeremy Silman who happened to be Stuart Cracraft’s chess teacher, and Stuart was nervous facing his own teacher. Of course, Deep Thought did not know who Jeremy Silman was. Jeremy managed to steer the game into an even position, but then lost in fifty-seven moves in a difficult ending.

深思团队。

The Deep Thought team.

从左到右:默里·坎贝尔、徐凤雄、托马斯·阿南塔拉曼、迈克·布朗和安德烈亚斯·诺瓦齐克。前景是 Deep Thought 板本身和 CDC 牌匾(超级计算机公司 CDC 为 ACM 锦标赛捐赠的牌匾)。拍摄于 1988 年《Deep Thought》因其大师级别的表演而获得弗雷德金中级奖之后拍摄的。

From left to right: Murray Campbell, Feng-hsiung Hsu, Thomas Anantharaman, Mike Browne, and Andreas Nowatzyk. In the foreground are the Deep Thought board itself and the CDC plaque (plaque donated by the supercomputer company CDC for the ACM Championship). Taken after Deep Thought won the Fredkin Intermediate Prize for its Grandmaster level performance in 1988.

与亚历克斯·菲什拜因(Alex Fishbein)的延期比赛是在两名球员都完成最后一轮比赛后进行的。Alex 输给了 GM Anthony Miles,后者在 8 分中以 6.5 分领先。 Deep Thought 是唯一可以赶上 Anthony 的球员。亚历克斯的绝招结果是失败的一招Deep Thought 很快就找到了获胜序列。这次斯图尔特没有宣布搭档,比赛进行到了八十,阿历克斯才认输。

The adjourned game against Alex Fishbein was played after both players had finished their last rounds. Alex lost his to GM Anthony Miles, who was leading the tournament with 6.5 points out of 8. Deep Thought was the only player that could catch up with Anthony. Alex’s sealed move turned out to be a losing one and Deep Thought quickly found the winning sequence. This time Stuart did not announce the mate, and the game was played out to move eighty before Alex conceded.

与肯·汤普森和大卫·斯雷特一起。

With Ken Thompson and David Slate.

左起:Ken Thompson(UNIX 图灵奖获得者、Belle 的作者之一)和 David Slate(Chess 4.X 的作者之一)。拍摄于 1989 年 WCCC(世界计算机国际象棋锦标赛)。

Standing from left: Ken Thompson (Turing Award winner for UNIX and one of the authors of Belle) and David Slate (one of the authors of Chess 4.X). Taken at the 1989 WCCC (World Computer Chess Championship).

克劳德·香农向作者展示了一个马头。

Claude Shannon presents the author with a horse’s head.

摄于 1989 年 WCCC。

Taken at the 1989 WCCC.

特级大师安东尼·迈尔斯 (Anthony Miles) 和深思 (Deep Thought) 共同获得了本次锦标赛的冠军,领先于包括前世界冠军米哈伊尔·塔尔 (Mikhail Tal) 在内的其他六位特级大师。这是第一次任何计算机在任何锦标赛中都领先于大师赛,更不用说前世界冠军了。由于 Deep Thought 没有资格获得奖金,安东尼·迈尔斯 (Anthony Miles) 获得了唯一的一等奖。

GM Anthony Miles and Deep Thought were the co-winners of the tournament, ahead of six other Grandmasters, including the former World Champion Mikhail Tal. This was the first time that any computer finished first in any tournament ahead of Grandmasters, let alone a former World Champion. Since Deep Thought was not eligible for prize money, Anthony Miles was awarded sole possession of first prize.

Fredkin中级奖的角逐正式结束。Deep Thought 在美国锦标赛中的表现为 2776,略低于名人堂成绩的 2790。1988 年,Deep Thought 参加了 42 场评级比赛,并且在所有 42 场比赛中的表现超过了 8 分,其中包括在美国公开赛上,有严重错误的比赛成绩为2598。在此期间最好的25场比赛成绩为2655,比之前高出155分Fredkin 中级奖的表现评级要求。

The race for the Fredkin Intermediate Prize was officially over. Deep Thought performed at 2776 on the US scale for the tournament, slightly behind the Hall of Fame result, which was 2790. In 1988, Deep Thought played forty-two rated games, and the performance over all forty-two8, including the ones played with serious bugs in the US Open, was 2598. The best twenty-five-game performance over the period was 2655 or 155 points higher than the performance rating requirement for the Fredkin Intermediate Prize.

Software Toolworks 比赛结束几个月后,Deep Thought 在加拿大埃德蒙顿每三年举办一次的世界计算机国际象棋锦标赛中凭借自己的实力成为世界冠军。然而,计算机国际象棋比赛对我们来说变得不再那么有趣。我们的全部注意力转向了击败人类世界国际象棋冠军的终极追求。

A few months after the Software Toolworks competition, Deep Thought became a World Champion in its own right in Edmonton, Canada, at the World Computer Chess Championship tournament that was held every three years. Computer chess competitions, however, were becoming less interesting to us. Our full attention turned to the ultimate quest of beating the human World Chess Champion.

最后一点,1989 年 Software Toolworks 锦标赛的组织者安排 Deep Thought 与 GM Anthony Miles 进行一场表演赛,作为 1988 年锦标赛的非官方决胜局。深思赢得了比赛。

As a final note, in 1989 the organizers of the Software Toolworks Championship arranged for Deep Thought to play GM Anthony Miles, in an exhibition game as the unofficial tiebreaker for the 1988 Championship. Deep Thought won the game.

    Berliner 博士是一名高级研究科学家,相当于正教授,但属于研究方向。

    Dr Berliner was a Senior Research Scientist, which was equivalent to a full professor but in a research track.

Monty Newborn 在其著作《卡斯帕罗夫 vs. 深蓝》    第一版中错误地指出 Hitech 于 1987 年成为高级大师。

    In the first edition of his book Kasparov vs. Deep Blue, Monty Newborn erroneously stated that Hitech became a senior master in 1987.

    对于不会下棋的人,这里有一个关于过路规则的解释。在国际象棋中,所有棋子都从棋盘的第二行开始。在文艺复兴之前,棋子一次只能向前移动一格。为了加快游戏速度,在文艺复兴时期,规则发生了变化,允许棋子从第二行开始的第一步向前移动两个方格。然而,这使得该棋子逃脱了被通常可以捕获第三排的棋子的相邻纵列上的对手棋子的捕获。添加了en passant(顺便)规则。在两格棋子移动之后,对手可以立即用相邻的棋子捕获刚刚移动的棋子,就像它是第三排的棋子一样。

    For people who do not play chess, here is an explanation of the en passant rule. In chess, all pawns start on the second rank of the board. Before Renaissance, pawns could only move one square forward at a time. To speed up the game, during the Renaissance, the rules were changed to allow the pawn to move two squares forward on its first move from the second rank. This, however, allowed the pawn to escape being captured by opponent’s pawn(s) on the adjacent files that normally could capture the pawn on the third rank. The en passant (in passing) rule was added. Immediately after a two-square pawn move, the opponent can, with an adjacent pawn, capture the pawn just moved as if it were a pawn on the third rank.

    Raj 后来成为卡内基梅隆大学计算机科学学院院长。

    Raj later became the Dean of the School of Computer Science at Carnegie Mellon University.

游戏得分和可能的组合    请参见附录 B。

    See Appendix B for the game score and the potential combination.

    开放对角线是指不再受到双方棋子阻碍的对角线。永久销钉是牢不可破的销钉,通常是在对手的主教将棋子钉在自己的王上时形成的。

    An open diagonal is a diagonal that is no longer impeded by pawns of either side. A permanent pin is an unbreakable pin, typically formed when a piece is pinned against its own king by an opponent’s bishop.

    到目前为止,Deep Thought 赢得了与 IMs 的所有比赛,除了对阵伊戈尔·伊万诺夫 (Igor Ivanov) 的一场平局之外。

    Up to this point, Deep Thought had won all its games against IMs, with the exception of one draw against Igor Ivanov.

    奇怪的是,42 在本章中反复出现,首先是 Deep Thought 对“生命、宇宙和一切”问题的回答,然后是 Bent Larsen 的世界排名,以及 Deep Thought 评分的游戏数量1988年的想法。

    It is a strange coincidence that 42 appeared repeatedly in this chapter, first as Deep Thought’s answer to the question of “Life, the Universe, and Everything”, then as the world ranking of Bent Larsen, and the number of rated games by Deep Thought in 1988.

 

 

第 6 章

CHAPTER 6

“敲门,敲门。谁在那儿?”

“Knock, Knock. Who’s There?”


喜欢和你做朋友的男人

The Man Who Likes to BE Your Friend

弗雷德里克·弗里德尔(Frederic Friedel)是一位德国自由记者,当时他不太忙于为他共同创立的两家与计算机国际象棋相关的公司工作。他的一家公司出版了一本计算机国际象棋杂志Computer Schach und Spiele,该杂志评论和推荐国际象棋计算机和国际象棋软件,而另一家公司 ChessBase 则生产国际象棋软件。存在潜在的严重利益冲突。但弗雷德里克并不是一个会因为利益冲突而止步的人,他对此持开放态度。

Frederic Friedel is a German freelance journalist, that is when he is not too busy working for the two computer chess-related companies that he co-founded. One of his companies publishes a computer chess magazine Computer Schach und Spiele, which reviews and recommends chess computers and chess software, while the other company, ChessBase, produces chess software. There is a potentially serious conflict of interest. But Frederic is not someone who would be stopped by a mere conflict of interest, and he is quite open about it.

安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 在汉堡大学时就认识了弗雷德里克。他通过一些活跃于电脑象棋的朋友认识了他几次。安德烈亚斯将弗雷德里克描述为开快车、穿着名牌西装、与名人交往的人。“和名人一起出去玩”这句话仍然是正确的,但我今天认识的弗雷德里克是一个有家庭的男人,有一个慈爱的妻子和两个聪明的孩子。

Andreas Nowatzyk knew Frederic from his time at the University of Hamburg. He had met him a couple of times through some friends who were active in computer chess. Andreas described Frederic as someone who drove fast cars, wore designer suits, and hung out with famous people. The “hung out with famous people” part is still correct, but the Frederic I know today is a family man with a loving wife and two bright kids.

弗雷德里克在德国境内组织了一些相当大型的计算机国际象棋比赛。其中之一是 1979 年 IM David Levy 和 Chess 4.8 之间的电视比赛。要么弗雷德里克很有说服力,要么大卫很勇敢。一个巨大的工业机器人在棋盘上进行了国际象棋 4.8 的走棋。如果有一两次电源闪烁,这个机器人很容易让大卫成为一个非常有名但受损的IM。

Frederic had organized some fairly large computer chess events on German soil. One of them was the 1979 TV match between IM David Levy and Chess 4.8. Either Frederic was very persuasive or David was very brave. A gigantic industrial robot made the moves for chess 4.8 over the chessboard. The robot could easily have made David a very famous but damaged IM, if there had been a power flicker or two.

弗雷德里克(Frederic)是一个友善的人,非常乐于助人,尤其是在相互帮助的情况下。由于弗雷德里克的两家公司都从事计算机国际象棋业务,因此他与许多顶级国际象棋棋手交往,通常为他们提供免费的国际象棋软件和数据库的国际象棋游戏。他很早就成为加里·卡斯帕罗夫的密友,甚至可能在加里被加冕为国际象棋世界冠军之前。据说加里在 ChessBase 本身的形成过程中发挥了重要作用。显然,ChessBase软件是一个国际象棋游戏数据库程序,是为了满足Garry的需要而创建的,该程序的许多功能都是基于Garry的规范。

Frederic was a friendly sort of guy who could be very helpful, especially when the help was mutual. Since both of Frederic’s companies are in computer chess, he socializes with many of the top chess players, usually offering them free chess software and databases of chess games. He became a close friend of Garry Kasparov very early, possibly even before Garry was crowned the World Chess Champion. Garry was said to be instrumental in the formation of ChessBase itself. Apparently, the ChessBase software, which is a database program for chess games, was created to satisfy Garry’s needs and many of the features of the program were based on Garry’s specifications.

1988 年 7 月下旬,安德烈亚斯在我们的办公室接到弗雷德里克打来的电话。随后,他向团队其他成员发送了一封电子邮件,总结了这次谈话。弗雷德里克邀请我们去德国参加一场表演赛。我很惊讶。尽管 ChipTest 赢得了 1987 年 ACM 锦标赛,但 Hitech 是更知名的机器。事实上,法国国际象棋杂志Europe Echecs在锦标赛结束后发表了一份报告,称 Hitech 以 2 比 1 的比例击败了 ChipTest。 Europe Echecs的报告可能错误引用或引用了错误的来源,因为 ChipTest 日志中的统计显示了一些东西完全不同。无论如何,在外界看来,海泰就是无冕之王。Fredkin Masters Open for Deep Thought 和 ChipTest 的结果打破了这一神话,但尚不清楚它们在卡内基梅隆大学之外是否广为人知。弗雷德里克是怎么知道我们的?

In late July 1988, Andreas received at our office a phone call from Frederic. Afterwards, he sent an e-mail to the rest of the team summarizing the conversation. Frederic was inviting us to play in an exhibition event in Germany. I was quite surprised. Although ChipTest had won the 1987 ACM Championship, Hitech was the better known machine. In fact, Europe Echecs, a French chess magazine had a report after the Championship that said Hitech was beating ChipTest by the ratio of 2 to 1. The Europe Echecs report probably misquoted or cited a wrong source, as the tally from ChipTest’s log showed something completely different. Anyway, to the outside world, Hitech was the uncrowned king. The results from the Fredkin Masters Open for Deep Thought and ChipTest dispelled the myth, but it was unclear as to whether they were well known outside of Carnegie Mellon. How did Frederic find out about us?

原来,弗雷德里克在贝尔全盛时期就和肯·汤普森成为了朋友。肯在澳大利亚悉尼贝尔实验室度过了一年的休假。由于 Ken 是我的论文委员会成员,我让他随时了解 ChipTest 和 Deep Thought 的最新进展,他显然也向 Frederic 介绍了我们。有时,认识合适的人是值得的。

It turned out that Frederic had befriended Ken Thompson back in Belle’s heyday. Ken was spending a sabbatical year from Bell Labs in Sydney, Australia. Since Ken was on my thesis committee, I kept him updated about what was happening with ChipTest and Deep Thought, and he apparently told Frederic about us. Sometimes, it pays to know the right people.

德国的邀请是德国最著名的计算机先驱康拉德·祖斯 (Conrad Zuse) 秋季生日庆祝活动的一部分。康拉德五十年前就曾预言,计算机将在五十年内击败国际象棋冠军。计划举办一场大型赛事,让最好的计算机对阵加里·卡斯帕罗夫。Frederic 正在组织它,并希望 Deep Thought 成为计算机。现在面对加里确实为时过早,但这是一个不容错过的好机会。无论如何,这将是一件令人兴奋的事情。

The German invitation was part of a birthday celebration in the fall for Conrad Zuse, Germany’s most prominent computer pioneer. Conrad had made a prediction fifty years earlier that a computer would defeat the chess champion within fifty years. A big event was planned to match the best computer against Garry Kasparov. Frederic was organizing it, and wanted Deep Thought to be the computer. It was definitely too early to face Garry but it was too good a chance to pass up. Anyway, it would be an exciting thing to do.

由于我们正忙于追逐弗莱德金中级奖,因此我们并没有真正跟踪邀请发生的情况。安德烈亚斯与弗雷德里克保持联系,并向团队其他成员通报情况。Frederic还主动提出提供Deep考虑到游戏数据库,我们很高兴地接受了这个提议。到 1988 年 9 月中旬,出现了一些新的进展。由于电视节目安排冲突,比赛被推迟到一月。另一个复杂因素是 IBM 德国是赞助商,并且极力要求在 IBM 主机上运行 Deep Thought。Deep Thought 当时无法在 IBM 主机上运行,​​因此比赛的资金存在疑问。一周后,我们被告知比赛取消。第一次尝试安排加里和深思之间的比赛就这样结束了。

Since we were busy chasing after the Fredkin Intermediate Prize, we did not really keep track of what was happening with the invitation. Andreas maintained contact with Frederic and kept the rest of the team informed. Frederic also offered to provide Deep Thought with game databases, and the offer was gladly accepted. By mid-September 1988, there were a few new developments. A TV scheduling conflict meant that the match was delayed until January. Another complication was that IBM Germany was the sponsor and was pressing hard to run Deep Thought on an IBM host computer. Deep Thought could not run on an IBM host at the time so the funding for the match was in doubt. A week later, we were informed that the match was off. Thus ended the first attempt to arrange a match between Garry and Deep Thought.


Frederic 多年来一直与团队保持联系。由于他也与加里关系密切,我们听到了冠军对国际象棋计算机的一些看法。以下摘录自他 1989 年写给我的一封信:

Frederic kept in touch with the team over the years. Since he was also close to Garry, we got to hear some of the Champion’s opinions about chess computers. The following excerpts came from a letter he sent me in 1989:

从 1988 年开始:“计算机永远无法击败大师。”

1988 年中:“计算机永远无法击败强大的大师。”

1988年底:“好吧,是的,本特是一位强大的大师。但计算机永远无法击败卡尔波夫或我。”

Beginning 1988: “Computers will never beat Grandmasters.”

Mid 1988: “Computers will never beat strong Grandmasters.”

End of 1988: “Okay, yes, Bent is a strong Grandmaster. But a computer will never be able to beat Karpov or me.”

默里·坎贝尔对最后的陈述有一个有趣的观察。默里在电脑国际象棋领域的时间比我长得多。多年来,他观察到,当计算机较弱时,国际象棋大师会说计算机永远无法击败国际象棋专家。当然,专家的演奏强度比大师低一级。随着时间的推移,同样的模式也出现在实力更强的玩家身上。当加里做出预测时,阿纳托利·卡尔波夫恰好是仅次于加里·卡斯帕罗夫的世界第二强选手。

Murray Campbell had an interesting observation in regard to this last statement. Murray had been in the computer chess field far longer than I had. Over the years, he had observed that when computers were much weaker, chess masters would say that computers would never beat chess experts. Of course, experts are one level below masters in playing strength. As time progressed, the same pattern repeated with players of greater strength. Anatoly Karpov happened to be the second strongest player of the world behind Garry Kasparov when Garry made his prediction.

IBM 连接

The IBM Connection

1988 年 9 月,IBM 德国取消了与加里的德国比赛。与此同时,美国的 IBM 研究中心已经与我们讨论了一段时间,讨论我们在 IBM 继续研究的可能性。据我所知,IBM德国公司并不知道我们毕业后可能会去IBM。

IBM Germany called off the German match with Garry in September 1988. Meanwhile, IBM Research in the US had been talking with us for some time about the possibility of continuing our research at IBM. As far as I know, IBM Germany had no idea that we might be moving to IBM after graduation.

IBM 兴趣的第一个暗示来自卡内基梅隆大学语音识别领域的研究生李开复。开复和一些 IBM 语音识别人员谈论了 Deep Thought 团队所做的事情,并得到了 IBM 可能有兴趣支持我们的反馈。研究。Peter Brown 是与开复交谈过的 IBM 人士之一。

The first hint of IBM interest came from Kai-Fu Lee, a Carnegie Mellon graduate student working in the speech recognition area. Kai-Fu talked to some IBM speech recognition people about what the Deep Thought team had done, and got the feedback that IBM might be interested in supporting our research. One of the IBM people that Kai-Fu talked to was Peter Brown.

彼得和我是卡内基梅隆大学 1982 年同一届新生。我们彼此认识,但并不是很亲密。弗雷德金大师公开赛前几周,我在计算机科学休息室遇到了彼得,并提到了我正在做的事情。Peter 问了很多问题,但我不知道他正在考虑让 IBM 雇用我们。

Peter and I had been in the same 1982 incoming class at Carnegie Mellon. We knew each other, but were not really close. I bumped into Peter in the Computer Science lounge a few weeks before the Fredkin Masters Open, and mentioned what I was doing. Peter was asking a lot of questions, but I had no idea that he was thinking about getting IBM to hire us.

多年后,Peter 描述了他如何让 IBM 的高层管理人员有兴趣在 IBM 内部继续该项目。当时正值超级碗赛季。Peter 碰巧和 IBM 计算机科学研究院副总裁 Abe Peled 一起在男洗手间。他们开始谈论超级碗电视广告的价格有多贵。彼得表示,他知道一种以更低的成本获得更大宣传的方法。他认识这群在卡内基梅隆大学研究计算机国际象棋的研究生,他相信这个团队将创造出第一台击败世界冠军的国际象棋机器。考虑到这一探索的历史意义以及潜在的公众利益,IBM 可以从这一努力中获得巨大的广告价值。哦,是的,该团队的成员都是世界级的人才,IBM 无论如何都会有兴趣雇用他们。安倍很感兴趣,并要求彼得调查一下。

Years later, Peter described how he got the upper management in IBM interested in continuing the project inside IBM. It was around the time of the Superbowl season. Peter happened to be in the men’s room with Abe Peled, the IBM Research Vice President of Computer Science. They started talking about how expensive the Superbowl TV commercial spots were. Peter suggested that he knew a way to gain much greater publicity at far lower cost. He knew this group of graduate students working on computer chess at Carnegie Mellon, and he believed that this team would create the first chess machine to beat the World Champion. Given the historical significance of the quest, and the latent public interest, IBM could stand to gain huge advertising value from the endeavor. Oh, yes, members of the team were world-class people that IBM would be interested in hiring anyway. Abe got interested and asked Peter to look into it.

约翰·科克 (John Cocke) 是现已退休的 IBM 研究员、图灵奖获得者,他率先提出了 RISC(精简指令集计算机)的理念,在团队聘用方面也发挥了重要作用。约翰是爱德华·弗雷德金教授的好朋友,爱德华·弗雷德金教授捐赠资金设立了计算机国际象棋弗雷德金奖。约翰对顶级计算机国际象棋团队来到 IBM 的可能性感到非常兴奋,他的兴奋具有感染力。到 1988 年 11 月初,几位 IBM 经理对此产生了兴趣,并要求我们提交申请。那是在我们在 Software Toolworks 锦标赛上引起轰动之前。

John Cocke, the now retired IBM fellow and Turing Award winner who pioneered the idea of RISC (Reduced Instruction Set Computer) among other things, was also instrumental in getting the team hired. John is a good friend of Professor Edward Fredkin, who donated the money to set up the Fredkin Prize for computer chess. John was quite excited about the possibility of the top computer chess team coming to IBM, and his excitement was contagious. By early November 1988, a couple of IBM managers were interested and we were asked to send in applications. This was before we made the splash at the Software Toolworks Championship.

Thomas、Murray 和我分别参观了 IBM 研究院并进行了推销。这个想法是将整个国际象棋机器集成到单个硅芯片上,然后使用大量芯片来构建“终极”国际象棋机器——一台可以击败世界冠军的国际象棋机器。我将负责芯片工作,而托马斯和默里将首先在中间机器上开发软件,然后将其移植到最终机器。到 1989 年 3 月上旬,Thomas 和我都获得了 IBM 研究中心的固定职位,从事国际象棋机方面的工作。默里出现了一些并发症,又花了一个月左右的时间才获得博士后职位。后来他的职位就变成了正规职位。

Thomas, Murray and I visited IBM Research separately and gave our sales pitch. The idea was to integrate an entire chess machine onto a single silicon chip, and then use a massive number of the chips to build the “ultimate” chess machine—a chess machine that could beat the World Champion. I would be the one doing the chip work, while Thomas and Murray would develop the software first on an intermediate machine, and then port it to the final machine. By early March 1989, both Thomas and I were offered regular positions at IBM Research to work on the chess machine. There were some complications with Murray, and it took another month or so before he was offered a post-doc position. His position later became a regular one.

在收到 IBM 通知的几周前,我收到了 Xerox PARC(帕洛阿尔托研究中心)计算机科学实验室经理的一封电子邮件,要求我考虑 PARC。施乐帕洛阿尔托研究中心 (Xerox PARC) 是计算机科学研究的传奇之地。20 世纪 70 年代和 80 年代一些最令人兴奋的计算机科学研究都来自那里。例如,Apple Macintosh 背后的主要思想可以追溯到 PARC 实验室。一些最具创新性的早期 VLSI 设计软件也来自 PARC。此外,实验室的氛围更接近我在卡内基梅隆大学习惯的氛围。PARC 也位于硅谷。我去过那里几次,真的很喜欢不同的氛围,更不用说天气了!硅谷距离台湾更近这一事实是一个额外的好处。我很快将我的简历发送给了帕洛阿尔托研究中心,并收到了积极的初步回复。

A few weeks before the notice from IBM, I received an e-mail from the manager of the Computer Science Laboratory at Xerox PARC (Palo Alto Research Center), asking me to consider PARC. Xerox PARC was a legendary place for computer science research. Some of the most exciting computer science research in the 1970s and 1980s came from there. For instance, the main idea behind the Apple Macintosh can be traced to the PARC lab. Some of the most innovative early VLSI design software also came from PARC. In addition, the lab atmosphere was much closer to what I was used to at Carnegie Mellon. PARC was also right in Silicon Valley. I had been there a few times, and really enjoyed the different atmosphere, not to mention the weather! The fact that Silicon Valley was much closer to Taiwan was a bonus. I quickly sent in my vita to PARC and received a positive initial response.

当 IBM 提供报价时,我仍在考虑将 PARC 作为替代方案。Thomas 和 Murray 都立即接受了 IBM 的报价,而我仍在等待 PARC 是否提供了更好的选择。我的处境相当尴尬。如果我拒绝 IBM 的邀请,Thomas 和 Murray 就需要在 IBM 寻找一个新项目,我不想让他们陷入困境,尤其是在我们一起经历过这些之后。但我也向帕洛阿尔托研究中心的人们保证要公平对待他们。此外,我喜欢去加利福尼亚的想法。“向西走吧,年轻人。” 我一直以为我最终会去硅谷。我的很多大学同学都在那里,还有我在卡内基梅隆大学的几个好朋友也往那个方向走。我决定推迟做出决定,直到我弄清楚 PARC 能提供什么。

When the IBM offer came in, I was still considering PARC as an alternative. Both Thomas and Murray accepted the IBM offer immediately, while I was still waiting to find out whether PARC provided a better option. I was in a fairly awkward position. If I were to decline the IBM offer, Thomas and Murray would need to find a new project at IBM, and I did not want to put them in a difficult situation, especially after what we had gone through together. But I had also promised the PARC people to give them a fair look. Furthermore, I liked the idea of going to California. “Go west, young man.” I always thought I would end up in Silicon Valley. A lot of my college classmates were there, and several of my good friends at Carnegie Mellon were also heading that way. I decided to delay the decision until I found out what PARC had to offer.

施乐帕克研究中心是一个非常有趣的地方,特别是对于在超大规模集成电路领域工作的人来说。不过,我确实有一些保留。施乐公司并非真正涉足计算机行业,它愿意长期支持这项工作吗?我也不知道如果施乐愿意支持该项目,我会做出什么决定,但只雇用了我们中的一个人。

Xerox PARC was a very interesting place, especially to someone who worked in the VLSI area. However, I did have some reservations. Would Xerox, which was not really in the computer business, be willing to support the effort long term? I also did not know what decision I would make if Xerox offered to support the project, but hired only one of us.

我内心的一部分欢迎再次独自工作的想法,至少是一段时间。团队一直以流水线方式工作,我先完成硬件,其他人随后完成软件。理论上,硬件和软件可以同时开发,但这只有在硬件定义良好的情况下才有可能,而新的国际象棋芯片肯定不是这样的。此外,新芯片的设计很容易需要几年的时间,而软件则可以在 6 个月到一年的时间内完成。IBM 聘用该团队的好处是可以保持团队的完整性,并为 IBM 提供即时的外部可见性。但就项目进度而言,收获相对较少;事实上,提供外部可见性的需要很容易延迟项目,因为我们随后必须构建一台中间机器以确保我们保持在计算机国际象棋世界的领先地位。独自工作时,我本来可以直接进入新的国际象棋芯片,这才是我真正想做的。

Part of me welcomed the idea of working alone again, at least for a while. The team had been working in an assembly line fashion, with me completing the hardware first, and the others finishing the software afterwards. Theoretically, the hardware and the software can be developed simultaneously, but that is only possible when the hardware is well defined, which was certainly not the case with the new chess chip. Also, the new chip could easily take a few years to design, while the software could be completed in say, 6 months to a year. IBM’s hiring of the team had the advantage of keeping the team intact, as well as providing IBM immediate external visibility. But as far as the project schedule was concerned, there was relatively little to be gained; in fact, the need to provide external visibility could easily delay the project, as we would then have to build an intermediate machine to ensure that we remain on top of the computer chess world. Working alone, I would have been able to go directly to the new chess chip, which was what I really wanted to work on.

不管怎样,我没有告诉帕洛阿尔托研究中心的人我会考虑单独工作的选择。也许,在潜意识里,我不想面对这个艰难的选择。我确实提到了 IBM 愿意提供什么。最后,PARC 决定他们无法匹配 IBM 聘请完整团队的提议,我通知 IBM Research 我接受了。

Anyway, I did not tell the people at PARC that I would consider the option of working alone. Perhaps, subconsciously, I did not want to face that difficult choice should it arise. I did mention what IBM was willing to offer. In the end, PARC decided that they could not match IBM’s offer to hire a full team, and I informed IBM Research of my acceptance.

DARPA 可怜的中校

The Poor Lieutenant Colonel at DARPA

1987年关于加州邀请的争论之后,我尽可能避免与柏林博士打交道。当然,他仍然是我的论文委员会成员,但考虑到我们之间的关系不太理想,破坏现状是没有意义的。艾伦·纽厄尔(Allen Newell)是该系的创始人之一,也是计算机象棋的先驱之一,他于 1989 年 4 月取代了 Berliner 博士在我的论文委员会中的地位。回想起来,我想知道 Berliner 博士离开委员会的原因是否是为了做更多的事情。与一名调查记者合作,而不是与该男子本人合作。

After the arguments over the invitation from California in 1987, I avoided dealing with Dr Berliner as much as possible. He was still on my thesis committee of course, but it made no sense to rock the boat given the less than ideal relationship we had. Allen Newell, who was one of the founders of the Department and also one of the pioneers of computer chess, replaced Dr Berliner on my thesis committee in April 1989. In retrospect I wondered if the reason for Dr Berliner leaving the committee were more to do with an investigative reporter than with the man himself.

Deep Thought 击败了 GM Bent Larsen,并在 1988 年 Software Toolworks 锦标赛中获得并列第一,这在国际象棋界是一个大新闻。在 Toolworks 锦标赛的同时,第 28 届国际象棋奥林匹克竞赛正在塞萨洛尼基举行,几乎所有世界顶级棋手都代表各自的国家队参赛。当 Deep Thought 在长滩获胜的消息传到塞萨洛尼基时,它成为了当天的热门话题。可能大多数顶级棋手在此之前从未听说过“深思”。

Deep Thought’s defeat of GM Bent Larsen, and tying for first in the 1988 Software Toolworks Championship, were big news in the chess world. At the same time of the Toolworks Championship, the twenty-eighth Chess Olympiad was being held in Thessaloniki, where nearly all of the world’s top chess players competed for their national teams. When the news of Deep Thought’s win in Long Beach was wired to Thessaloniki, it became the talk of the day. Probably most of the top chess players had never heard of Deep Thought until then.

过了一段时间,大众媒体才开始关注这个消息。12 月 22 日,也就是比赛结束一个月后, 《国际先驱论坛报》的头版刊登了这样的标题:“计算机优先:‘深度思考’震撼国际象棋世界”。我们团队中没有人知道我们已经登上了一份国际报纸的头版。据我们所知,我们没有登上美国任何全国性报纸的头版,当然我们也不希望看到这条新闻出现在国际报纸上。这个标题引起了伦敦一名记者的注意,并决定展开“调查”。

It took a while before the general press picked up on the news. On December 22, a month after the tournament, the front page of the International Herald Tribune carried the headline “A Computer First: ‘Deep Thought’ Stuns the Chess World”. None of us on the team had any idea that we had made the front page of an international newspaper. As far as we knew, we did not make the front page of any national newspaper in the US, and certainly we were not expecting to see the news showing up in an international newspaper. This headline caught the attention of a reporter in London who decided to start an “investigation”.

我们从 Berliner 博士那里得知了记者对该项目的兴趣,他打电话给我们,说一位记者有兴趣玩 Deep Thought 的演示游戏。柏林纳博士补充说,这位记者来自一家备受尊敬的杂志社。由于我们开始寻找工作并且可以利用宣传,我们欣然同意设置演示。在约定的时间,记者打来电话,介绍自己是英国《观察家》杂志的多米尼克·劳森。后来我才知道,他是该杂志的副主编(现在是英国《星期日电讯报》的编辑)。多米尼克输掉了比赛,我们以为故事就这样结束了。他没有尝试采访 Deep Thought 团队的任何成员,似乎只是对玩这个游戏感兴趣。

We learned about the reporter’s interest in the project from Dr Berliner, who called us to say that a reporter was interested in playing a demo game with Deep Thought. Dr Berliner added that the reporter was from a well-respected magazine. Since we were starting to look for jobs and could use the publicity, we readily agreed to set up the demo. At the agreed time, the reporter called and introduced himself as Dominic Lawson from the British magazine The Spectator. I found out later that he was the Deputy Editor of the magazine (he is now Editor of The Sunday Telegraph newspaper in Britain). Dominic lost the game and we thought that was the end of the story. He made no attempt to interview any member of the Deep Thought team and seemed only to be interested in playing the game.

然后在 1989 年 2 月 27 日,有人给我看了查尔斯·克劳萨默 (Charles Krauthammer) 的专栏,该专栏由《华盛顿邮报》联合发表在《匹兹堡邮报》上。这是一篇写得非常好的文章,介绍了 Deep Thought 在长滩的成功及其影响,并引用了《旁观者》作为部分材料的来源。

Then on 27 February 1989, someone showed me the column of Charles Krauthammer, syndicated from the Washington Post, in the Pittsburgh Post-Gazette. It was a very well written article about Deep Thought’s success in Long Beach and its implications, and quoted The Spectator as the source for some of the material.

卡内基梅隆大学的一个图书馆订阅了《旁观者》杂志,我发现了查尔斯·克劳萨默引用的这篇文章。这不仅仅是一篇文章,而是该期的封面故事。封面上,一架机身呈棋子象形状、四台发动机呈棋子形状的飞机正在方格田野上空飞行。一面美国国旗显眼地展示着。封面标题是“为胜利而战”,较小的标题是“多米尼克·劳森揭示了五角大楼的战争举动”。我皱起眉头,“安德烈亚斯不会喜欢这样的。” 安德烈亚斯对德国绿党持同情态度,该党热衷于环境事业和和平主义。支持任何形式的战争努力都将完全违背他的信念。就此而言,团队中的其他成员都没有对所建立的联系感到兴奋。

One of the libraries at Carnegie Mellon had a subscription to The Spectator and I found the article quoted by Charles Krauthammer. It was not just an article, but the cover story of that issue. On the cover, an airplane, with the fuselage in the shape of the chess piece bishop and four engines in the shape of chess pawns, was flying over a checkered field. An American flag was prominently displayed. The cover headline read “Playing to Win” and the smaller headline read “Dominic Lawson reveals the Pentagon’s war moves”. I frowned, “Andreas is not going to like this.” Andreas had a sympathetic view for Germany’s Green Party, which was into environmental causes and pacifism. Supporting any sort of war effort would be completely counter to his beliefs. For that matter, none of the other members of the team was thrilled about the connections that were made.

从某种意义上说,多米尼克的文章很搞笑。《旁观者》似乎是一本认真对待自己的杂志,但这篇文章在我看来毫无根据。杂志内的副标题是“多米尼克·劳森调查“深思”项目,发现五角大楼正在利用国际象棋进行军备竞赛”。好吧,我不知道——现在仍然不知道——五角大楼是否正在利用国际象棋来进行军备竞赛。但我确实知道“深度思考”并不是因为五角大楼的某人想要完成而开始的。

In one sense, Dominic’s article was hilarious. The Spectator appeared to be a magazine that took itself seriously, but the article seemed to me to be pretty groundless. The subtitle inside the magazine read “Dominic Lawson investigates the Deep Thought project, and discovers that the Pentagon is harnessing chess to the arms race”. Okay, I did not know—and I still do not know—whether the Pentagon was harnessing chess to the arms race. But I do know that Deep Thought was not started because someone in the Pentagon wanted it done.

当我开始阅读这篇文章时,事实错误不断出现,而且这只是我确定有错误的部分。文章的一部分写道:“卡内基梅隆大学的 Deep Thought 项目和其他计算机国际象棋开发项目的全部费用均由五角大楼国防高级研究计划局 DARPA 支付”。所以,我们从学生口袋里掏出来的钱也属于 DARPA!

Once I started reading the article, factual errors kept on popping up, and that was just for the parts that I knew for sure were wrong. One section of the article read, “The Deep Thought project, and other computer chess developments at Carnegie-Mellon, have been paid for in their entirety by DARPA, the Pentagon’s Defense Advanced Research Projects Agency”. So, the money we paid out of our student pockets belonged to DARPA as well!

然后我读到了这篇文章中最有趣的部分。“我请 DARPA 专家系统研究(包括 Deep Thought 项目)资金负责人 Robert Simpson 中校给我一个关于 Deep Thought 的明确的军事应用。” 首先,Deep Thought并不是一个“专家系统”,这在人工智能研究领域是非常具体的。该团队的一些成员可能会认为将 Deep Thought 称为专家系统是一种侮辱。事实上,团队中至少有两名成员使用了“废话”这个词来描述“专家系统”,或者就此而言,人工智能。其次,Deep Thought 研究部分得到了 VLSI 项目的支持,并且没有任何资金来自 DARPA 对专家系统研究的资助。据我所知,罗伯特·辛普森中校从未收到过任何描述“深思”项目的纸片。

Then I read what was the most amusing part of the article. “I asked Lieutenant Colonel Robert Simpson, DARPA’s head of funding for expert systems research—which includes the Deep Thought project—to give me one clear military application of Deep Thought.” First, Deep Thought was not an “expert system”, which means something quite specific in the research field of Artificial Intelligence. Several members of the team would probably consider it an insult to call Deep Thought an expert system. In fact, at least two members of the team had used the word bullshit to describe “expert systems”, or for that matter, Artificial Intelligence. Second, the Deep Thought research was supported in part by the VLSI project, and none of the money came from the DARPA funding for expert systems research. Lieutenant Colonel Robert Simpson had to my knowledge never received a single scrap of paper describing the Deep Thought project.

想象一下这位中校所处的情况,记者打电话给你,询问你一个你一无所知的项目。你必须说点什么。嗯,罗伯特·辛普森中校头脑敏捷。

Just imagine the situation that this Lieutenant Colonel was placed in. A reporter calls you up and asks you about a project that you know nothing about. You have to say something. Well, Lieutenant Colonel Robert Simpson had a quick mind.

“‘在战场情况下导航,’他立即回答道。“像这样的机器,根据飞行员正在飞行的地形知识进行编程,可以将所有不同的路线选择数字化,探索它们,并选择最佳路线。这正是 Deep Thought 正在做的事情。”有趣的是,我没想到卡斯帕罗夫先生会制造出一枚不错的巡航导弹。

“‘Navigation in a battlefield situation,’ he replied promptly. ‘A machine like this, programmed with knowledge of the terrain a pilot is flying through, can digitize all the various route choices, explore them, and choose the optimum route. That is exactly what Deep Thought is doing.’” Interesting, I hadn’t thought that Mr. Kasparov would make a nice cruise missile.

然后我遇到了一些非常烦人的事情。“Deep Thought 项目的负责人兼卡内基梅隆大学计算机科学系的 Hans Berliner 教授警告我,大约三年后,他的团队将产生一个 Deep Thought 的儿子,它的国际象棋水平将比世界冠军还要强。 ,加里·卡斯帕罗夫。“这真的很遗憾,”他叹了口气。“就像国际象棋棋手开始过上体面的生活一样。”柏林纳博士既不是“深思”项目的负责人,也不是部门的负责人,他是一个竞争对手。多米尼克对柏林博士的采访到底发生了什么,柏林博士与深度思考完全没有关系?

Then I came to something quite annoying. “The head of the Deep Thought project and of the Carnegie-Mellon computer science department, Professor Hans Berliner, warned me that in about three years his team would have produced a son of Deep Thought, which would be stronger at chess than the world champion, Garry Kasparov. ‘It’s a pity, really,’ he sighed. ‘Just as chess players were beginning to make a decent living.’” Dr Berliner was neither the head of the Deep Thought project nor the head of the department, and he was a competitor. What had gone on in Dominic’s interview with Dr Berliner, who had had absolutely nothing to do with Deep Thought?

当我平静下来时,我意识到现实可能是多米尼克心里有一个好故事,但可能没有时间检查所有细节。毕竟,在其他媒体对 Deep Thought 的报道中,柏林纳博士被错误地称为该项目的领导者。

When I calmed down, I realized that the reality might be that Dominic had a good story in mind, and may not have had the time to check all the details. Dr Berliner had after all erroneously been named as leader of the project in other media coverage of Deep Thought.

不管怎样,在团队的其他成员看到这篇文章后,我们决定给《旁观者》寄一封信。我还给柏林博士发了一封电子邮件,请他确保其他记者不会犯同样的错误。我们仍在寻找工作,错误的归因可能会影响我们的机会。一件事接二连三,一个月后,艾伦·纽厄尔教授取代了我论文委员会的柏林纳博士。

Anyway, after the rest of the team had seen the article, we decided to send a letter to The Spectator. I also sent an e-mail to Dr Berliner asking him to please make sure that other reporters would not make the same mistake. We were still looking for jobs, and the wrong attribution could affect our chances. One thing led to another, and a month later, Professor Allen Newell replaced Dr Berliner on my thesis committee.

在我们将这封信寄给《旁观者》几周后,一封经过大规模修改的信出现在他们的信件部分。

A few weeks after we mailed the letter to The Spectator, a massively modified letter appeared in their letters section.

先生:《五角大楼下棋》(1 月 28 日)这篇文章包含一些严重的误解和事实错误。

Sir: The article, ‘The Pentagon Plays Chess’ (28 January), contains several gross misconceptions and factual errors.

(这和我们实际写的不一样,但本质是一样的。)

(This is different from what we actually wrote, but the essence is the same.)

Deep Thought 项目并不完全由 DARPA、国防部或五角大楼直接资助。这些支持包括工业资金、国家科学基金会资金,以及某种程度上的国防部资金。

The Deep Thought project is not entirely funded by DARPA, the Department of Defense, or the Pentagon directly. The support includes industrial funding, National Science Foundation funds, and, to some degree, Department of Defense money.

这与我们写的内容截然不同,“深度思考项目不是由 DARPA/DOD 或五角大楼直接资助的,当然也不是全部资助的。Deep Thought 项目是一项娱乐练习,几名研究生应用了他们所学到的有关 VLSI 电路和硬件的知识设计。第一个结果是 ChipTest,它完全是我们利用业余时间其他项目的废料构建的。ChipTest 在 1987 年 ACM 计算机国际象棋锦标赛中取得成功后,我们的顾问之一 HT Kung 教授从他的自由支配基金中拿出了 5000 美元,帮助我们构建了当前的 Deep Thought。这笔钱与任何官方项目没有直接关系,而是典型大学正常运营费用的一部分。因此,它包括工业资金、学费、NSF(国家科学基金会)资金、捐款,以及某种程度上的国防部资金。此外,Deep Thought 的一些各种费用都是我们自己掏腰包的。”

This is drastically different from what we wrote, which was “The Deep Thought project is not funded by DARPA/DOD, or the Pentagon directly, and certainly not in its entirety. The Deep Thought effort was a recreational exercise where several graduate students applied what they learned about VLSI circuit and hardware design. The first result was ChipTest, which we built entirely out of scraps from other projects in spare time. After ChipTest’s success in the 1987 ACM Computer Chess Championship, one of our advisors, Prof. H. T. Kung, helped us with a modest $5000 out of his discretionary fund, which was used to construct the current Deep Thought. This money is not directly related to any official project, but is part of normal operational expenses of typical universities. As such, it includes industrial funding, tuition, NSF (National Science Foundation) funds, donations, and to some degree, DOD money. In addition, some of the various expenses for Deep Thought came right out of our own pockets.”

公开的信继续说道。

The published letter continued.

我们非常怀疑罗伯特·辛普森中校是否意识到深度思考的本质,这与专家系统研究完全无关。Deep Thought 是一款专门用于下棋且仅用于国际象棋的引擎。Deep Thought 使用的主要算法几十年前就已为人所知,而且 Deep Thought 的结构对于战场导航没有任何用处。引用贝尔实验室的肯·汤普森的话说,“Belle(像 Deep Thought 这样的国际象棋机器)唯一的军事应用就是从飞机上扔下它并杀死某人。”

We very much doubt that Lieutenant Colonel Robert Simpson was aware of the nature of Deep Thought which is completely unrelated to expert systems research. Deep Thought is a special purpose engine for playing chess and chess only. The principal algorithms used by Deep Thought had been known for decades, and the structure of Deep Thought has no utility for battlefield navigation whatsoever. To quote Ken Thompson of Bell Labs, “the only military application for Belle [a chess machine like Deep Thought] is to drop it from an airplane and kill someone.”

本段已完整给出。

This paragraph was given in its entirety.

以下内容并未出现在《旁观者》中。“Hans Berliner 博士是卡耐基梅隆大学计算机科学系的高级研究人员,负责 Hitech 国际象棋机的研究项目。他不是计算机科学系的系主任(即 Nico Habermann 教授),也没有以任何方式参与 Deep Thought 的工作。

The following did not show up in The Spectator. “Dr Hans Berliner is a senior research faculty member of the CMU Computer Science Department, and heads the research project on the Hitech chess machine. He is not head of the Computer Science Department, who is Professor Nico Habermann, nor is he involved with the Deep Thought effort in any way.

劳森先生在对 Deep Thought 项目的“调查”中,完成了一项不寻常的壮举,没有采访 Deep Thought 团队的任何成员。

Mr. Lawson, in his ‘investigation’ of the Deep Thought project, managed the unusual feat of interviewing not a single member of the Deep Thought team.”

这封信一定太长了。

The letter must have been too long.

从国际象棋到华尔街

From Chess to Wall Street

当我决定追逐国际象棋世界冠军时,只要有机会,我就会在《纽约时报》上阅读大师罗伯特·伯恩的专栏。当 ChipTest 投入运行时,我会使用 ChipTest 检查 Robert 的分析。有时 ChipTest 会提出他错过的想法,但他经常会进行一些 ChipTest 很难复制的深入分析。

Once I decided to go after the World Chess Champion, I read Grandmaster Robert Byrne’s columns in the New York Times whenever I could lay my hands on them. When ChipTest became operational, I would check out Robert’s analysis with ChipTest. Sometimes ChipTest would come up with ideas that he missed, but frequently he would have some deep analysis that ChipTest had a hard time replicating.

当 ChipTest 赢得 1987 年 ACM 冠军时,Robert 写了一篇相关文章,但直到 1988 年 ACM 冠军 Deep Thought 击败 Hitech 后我们才真正接触到他。赛事总监迈克·瓦尔沃 (Mike Valvo) 向罗伯特提供了有关 Deep Thought 的背景信息。由于很多人没有意识到 Berliner 博士与 Deep Thought 无关,我们请 Mike Valvo 帮助澄清情况。这是我们第一次与罗伯特间接接触。

When ChipTest won the 1987 ACM Championship, Robert wrote an article about it but we did not really have contact with him until after the 1988 ACM Championship where Deep Thought defeated Hitech. Mike Valvo, the tournament director, was providing Robert with background information about Deep Thought. Since many people did not realize that Dr Berliner was unrelated to Deep Thought, we asked Mike Valvo to help clarify the situation. It was our first indirect contact with Robert.

虽然 Deep Thought 从来都不是一个人工智能项目,但人工智能界将 Deep Thought 视为自己的项目,我们受邀在 1989 年国际人工智能联合会议上与总经理罗伯特·伯恩 (Robert Byrne) 进行了一场表演赛。两名球员都不在现场。Deep Thought 在匹兹堡比赛,而 Robert 在纽约州的家中比赛。这些举动通过今日美国体育中心转播,这是一个类似于 AOL 或 MSN 的在线服务。深思赢得了比赛,但几天后罗伯特在体育中心安排的第二场比赛中还以颜色。罗伯特和深思又打了两场比赛,各赢一场。

While Deep Thought was never an Artificial Intelligence project, the AI community treated Deep Thought as one of their own, and we were invited to play an exhibition match with none other than GM Robert Byrne at the 1989 International Joint Conference on Artificial Intelligence. Neither player was on site. Deep Thought played from Pittsburgh, while Robert played from his home in New York State. The moves were relayed over the USA Today Sports Center, an on-line service similar to AOL or MSN today. Deep Thought won the game, but a few days later Robert returned the favor in a second game arranged by the Sports Center. Robert and Deep Thought played two more games with one win each.

在这四场比赛之后,《纽约时报》对这个项目产生了兴趣,罗伯特被要求为该论文的《科学时报》版块写一篇文章。他为了这篇文章访问了卡内基梅隆大学,这是我们第一次面对面见到他。他看上去就像一位尊贵的老绅士,人们在认识他时最常用的形容词就是“一位和蔼可亲的绅士”。我们立刻就喜欢上了他。我们加入IBM后,发现罗伯特就住在附近。穆雷也是一位狂热的网球运动员(在穆雷的女儿出生之前),他们经常在周末打球。罗伯特偶尔会过来用我们最新的机器玩一些游戏。除了参与该项目的大师之外,罗伯特可能比任何其他大师都更了解我们的机器。

After these four games, the New York Times took an interest in the project, and Robert was asked to write an article for the Science Times section of the paper. He visited Carnegie Mellon for the article, and it was the first time we met him face to face. He looked like a distinguished old gentleman—the most frequent description used when people got to know him was “a sweet gentleman”. We all liked him immediately. After we joined IBM, we found out that Robert was living close by. An avid tennis player as was Murray (before Murray’s daughter was born), they would frequently play on weekends. Once in a while, Robert would drop in and play some games with our latest machine. Except for the Grandmasters who worked on the project, Robert probably knew more about our machines than any other Grandmaster.

罗伯特在《科学时报》上发表的文章是封面故事,长达两页。其中描述了 Andreas 在评估函数自动调整方面的工作。华尔街一家股票和商品交易公司的首席执行官看到这篇文章后给安德烈亚斯打电话,为他提供了一份高薪工作。出于好奇,安德烈亚斯去了纽约,结识了华尔街人士。想要给人留下深刻印象,首席执行官带他去了证券交易所,但安德烈亚斯穿着牛仔裤,而且不允许穿牛仔裤上场。首席执行官不得不说服保安看向别处。接下来的楼层参观本来只是为了一睹为快,但不知何故却变成了更详细的参观。监控摄像头拍到安德烈亚斯在交易大厅穿着牛仔裤,保安人员下楼,随后发生了争执。我不了解安德烈亚斯,但如果我处在他的位置,一旦我被从地板上清理出来,我就会开怀大笑。华尔街是一个不同的世界。

Robert’s article in the Science Times was the cover story, spanning two pages. In it, Andreas’ work on the automatic tuning of the evaluation function was described. A CEO of a Wall Street stock and commodity trading firm saw the article and gave Andreas a call, offering him a high paying job. Out of curiosity, Andreas went to New York and met the Wall Street people. Wanting to impress, the CEO took him to the stock exchange, but Andreas was wearing jeans, and jeans were not allowed on the floor. The CEO had to talk the security guard into looking the other way. The following floor trip was supposed to be for a peek only, but somehow it became a more detailed tour. The security cameras caught Andreas wearing jeans on the exchange floor, the security people descended, and an argument ensued. I don’t know about Andreas, but if I were in his place, I would have had a hearty laugh once I was cleared from the floor. Wall Street was a different world.

虽然华尔街的提议很有趣,但安德烈亚斯拒绝了。这家华尔街公司继续提高报价,但安德烈亚斯仍然不感兴趣。

While the Wall Street proposal was interesting, Andreas turned it down. The Wall Street firm continued to sweeten the offer, but Andreas remained uninterested.

1990 年 2 月,Andreas 接受了加利福尼亚州 Sun Microsystems 的一份工作,入职日期为 1990 年 6 月。他还有一些论文需要撰写。此时,默里和我都已经前往 IBM。华尔街公司邀请安德烈亚斯在他完成论文期间做一些咨询工作。安德烈亚斯将咨询工作视为写论文后的带薪休假,因此去了纽约四个月,偶尔回匹兹堡。当时,这家华尔街公司使用的软件状况相当糟糕。它确实有效,但速度就像蜗牛一样。对于 Andreas 正在开发的软件,数据的总存储量约为 10 MB。软件总是回到磁盘获取数据,尽管磁盘访问速度比主存访问速度慢一千倍以上。安德烈亚斯重写了软件,将所有数据读入主内存,瞧,它现在的运行速度快了一千倍以上。华尔街人士对此印象深刻。安德烈亚斯还做了一些在他们眼中令人惊奇的事情。对他们来说,安德烈亚斯就像一个可以在水上行走的人。如今,许多真正的计算机科学家在华尔街工作,安德烈亚斯发现的问题不太可能再次出现。

In February 1990, Andreas accepted a job at Sun Microsystems in California, with the starting date in June 1990. He still had some writing to do for his thesis. By this time both Murray and I had already left for IBM. The Wall Street firm came back with an offer for Andreas to do some consulting work while he was finishing up the thesis. Thinking of the consulting work as a paid vacation from writing his thesis, Andreas went to New York for four months, with occasional trips back to Pittsburgh. At the time, the software used in the Wall Street firm was in pretty bad shape. It worked, but at the speed of a snail. For the piece of software that Andreas was working on, the total storage for the data was something like 10 megabytes. The software always went back to the disk for the data, even though the disk access speed was more than a thousand times slower than the main memory access speed. Andreas rewrote the software to read in all the data into main memory, and voilà, it now ran more than a thousand times faster. The Wall Street people were impressed. Andreas did some other amazing things in their eyes. To them, Andreas was like someone who could walk on water. Today, many real computer scientists work on Wall Street and the problem that Andreas found is unlikely to reappear.

安德烈亚斯到Sun Microsystems工作后,这家华尔街公司拒绝放弃他,并向他提出了一个他们认为他无法拒绝的提议。该提案是这样的:该公司将在硅谷创建一家拥有巨额年度预算的公司。安德烈亚斯将领导公司,并对公司将做什么有充分的发言权,但他一半的时间必须花在继续他为公司所做的事情上。安德烈亚斯将可以自由雇用人员,这些人员也将获得利润分享协议等。安德烈亚斯问托马斯·阿南塔拉曼迈克·布朗 (Mike Browne) 以及工程实验室的高级工程师约翰·扎尔奈 (John Zsarnay),无论他们是否感兴趣。他们是。约翰去了加利福尼亚州并开始建立基础设施。但随后安德烈亚斯又改变了主意并退出了这笔交易。该公司与托马斯和迈克重新谈判了交易,约翰被搬回东海岸。约翰很快离开并加入了当时硅谷的一家初创公司 Rambus。直到很久以后我才意识到这一切,但整个事件最终对 IBM 的项目产生了意想不到的影响。

After Andreas went to work at Sun Microsystems, the Wall Street firm refused to give up on him and offered him a proposal that they thought he could not refuse. The proposal went something like this: the firm would create a company in Silicon Valley with a hefty annual budget. Andreas would head the company and have full say on what the company would do, but half of his time had to be spent on continuing what he was doing for the firm. Andreas would be free to hire people who would also be given profit sharing deals and so on. Andreas asked Thomas Anantharaman and Mike Browne, as well as John Zsarnay, a staff engineer at the Engineering Lab, whether they were interested. They were. John went to California and started setting up the infrastructure. But then Andreas had second thoughts and backed out of the deal. The firm renegotiated the deals with Thomas and Mike, and John was moved back to the East Coast. John soon left and joined Rambus, a startup at the time in Silicon Valley. I was quite unaware of all this until much later, but the whole episode eventually would have an unexpected influence on the project at IBM.


Deep Thought 团队中的任何人都没有想到我们的工作会与华尔街有某种联系。该项目的自动调整部分本身就是一个有趣的话题。但这么做的主要原因还是来自于加州邀请事件。谁会想到加利福尼亚州的国际象棋锦标赛邀请会导致我们中的一些人为一家华尔街公司工作?有时,生活充满了最奇怪的曲折。

None of us on the Deep Thought team ever imagined that our work would have some link with Wall Street. The automatic tuning part of the project was an interesting topic in its own right. But the main reason for doing it came from the incident with the California invitation. Who would have thought that the invitation from California for a chess tournament would lead to some of us working for a Wall Street firm? Sometimes, life is full of the strangest twists and turns.

 

 

间奏曲

Intermezzo


 

 

第7章

CHAPTER 7

与历史的第一次约会

First Date with History


从纽约打来的电话

A Phone Call from New York

当Deep Thought在国际象棋界出名后,各种电话开始打来。通常,他们会被转接到我们的办公室。然而,有一天晚上,我在家接到一个我从未听说过的人打来的电话,他问了很多问题。据说他是从明尼苏达州的某个地方打来的,与任何新闻机构都没有关系,纯粹是出于个人兴趣打来的。显然,他在白页上找到了我的电话号码。他看起来无害,但与他交谈我确实感到紧张。在我离开卡内基梅隆大学前往 IBM 之前的几个月里,他每隔几周就会打电话询问 Deep Thought 的最新动态。

After Deep Thought became famous in the chess world, phone calls started to come in from all sort of places. Usually, they were put through to our offices. One night, however, I got a call at home from some guy I had never heard of, who was asking lots of questions. He was supposedly calling from somewhere in Minnesota, was not associated with any news organization, but was calling purely out of personal interest. Apparently, he found my phone number in the white page. He seemed harmless, but I did get nervous talking to him. For several months until I left Carnegie Mellon for IBM, he would call every few weeks trying to find out what was new with Deep Thought.

因此,当我在家中又接到一个意外的电话时,我有点紧张。虽然那声音确实有些熟悉,但我确信那不是我认识的人。“不是另一个,”我想。来电者介绍自己是谢尔比·莱曼。

I was therefore a little bit nervous when I got another unexpected phone call at home. Although the voice did seem somewhat familiar, I was sure that it was not someone I knew. “Not another one,” I thought. The caller introduced himself as Shelby Lyman.

在 1972 年鲍比·费舍尔 (Bobby Fischer) 和鲍里斯·斯帕斯基 (Boris Spassky) 之间的“世纪之战”之前,国际象棋在美国还没有电视转播。谢尔比·莱曼和他的朋友说服 PBS(公共广播系统)对这场比赛制作电视节目,由谢尔比担任主要评论员。一时间,随着费舍尔与斯帕斯基的比赛吸引了街上人们的注意力,谢尔比立即成为名人。在鲍比·费舍尔(Bobby Fischer)拒绝按照国际棋联的条件与俄罗斯挑战者阿纳托利·卡尔波夫(Anatoly Karpov)捍卫自己的冠军头衔后,公众对国际象棋的兴趣减弱了。然而谢尔比坚称与 PBS 建立了良好的关系,并能够让 PBS 播放更多报道世界象棋锦标赛的电视节目。

Before “the match of the century1” between Bobby Fischer and Boris Spassky in 1972, chess was not telecast in the USA. Shelby Lyman and his friends persuaded PBS (Public Broadcasting Systems) to do a TV show on the match, with Shelby being the main commentator. For a while, with the Fischer-Spassky match grabbing the attention of people on the streets, Shelby was an instant celebrity. The public’s interest in chess waned after Bobby Fischer refused to defend his title under FIDE’s conditions against his Russian challenger Anatoly Karpov. Shelby, however, maintained a good relationship with PBS and was able to get PBS to broadcast more TV shows covering the World Chess Championships.

在决定要击败世界冠军后,我开始观看谢尔比节目的一些剧集。说实话,谢尔比的表演有点慢,而且太做作,不合我的口味。该节目在 1972 年可能会很受欢迎,当时人们对费舍尔与斯帕斯基的比赛兴趣很高,但随着 MTV 和电子游戏的出现,该节目根本没有引起共鸣。

I started viewing some episodes of Shelby’s show after deciding that I wanted to beat the World Champion. To tell the truth, Shelby’s show was a little bit slow and too artificial for my taste. The show might have worked in 1972 when the interest in the Fischer vs. Spassky match was high, but with the MTV and video game generation, the show simply did not resonate.

谢尔比问我们是否有兴趣与国际象棋世界冠军加里·卡斯帕罗夫进行一场表演赛。当然,答案是肯定的。我们还没有真正准备好与世界冠军比赛,但我们不想错过这个机会。我确实有些怀疑谢尔比能否让比赛顺利进行。尽管潜在组织者与加里关系密切,但德国早先提出的表演赛提案最终被取消。

Shelby asked whether we would be interested in playing an exhibition match with Garry Kasparov, the World Chess Champion. Of course, the answer was yes. We were not really ready to play the world champion, but we did not want to miss the chance. I did have some doubt whether Shelby could make the match happen. The earlier exhibition match proposal from Germany was scrapped in the end, despite the close relationship between the potential organizer and Garry.

当谢尔比打电话时,我正在设计 Deep Thought 的印刷电路板版本。卡内基梅隆大学的 Raj Reddy 教授对 Deep Thought 项目很感兴趣,并提出为我们提供大约 10,000 美元,为带有两个以上国际象棋处理器的 Deep Thought 版本制作额外的板卡。元件成本已经充分降低,我们可以构建四块印刷电路板,每块板上有两个国际象棋处理器。理论上,对于拟议的比赛,我们可以让一台新机器运行速度达到 Deep Thought 的五倍(四块新板加一块现有板)。

At the time that Shelby called, I was designing a printed circuit board version of Deep Thought. Professor Raj Reddy at Carnegie Mellon had taken an interest in the Deep Thought project, and offered to provide us with about $10,000 to make additional boards for a version of Deep Thought with more than two chess processors. The component cost had come down sufficiently that we could build four printed circuit boards with two chess processors on each board. In theory, for the proposed match we could have a new machine running that was up to five times as fast as Deep Thought (four new boards plus one existing board).

由于对印刷电路板的机械设计了解不多,我犯了一个严重的错误。深思板有一个大披萨那么大。对于这种尺寸的电路板,焊接过程中的热膨胀会导致电路板严重翘曲。虽然我确实在新的印刷电路板上放置了一个板加强筋以防止正常处理过程中的翘曲,但它的强度不足以防止焊接过程中产生的翘曲。所有四块板都明显变形。可能是由于翘曲,四块新板中的一块从未工作过。也许需要做大量的工作才能修复损坏的主板,但在我前往 IBM 之前,我有一篇论文要答辩,还有很多行李要做。因此,我们这场比赛真正能拥有的最大板数是四块。

Not knowing much about the mechanical design of printed circuit boards, I made a serious error. The Deep Thought boards were the size of a large pizza. For circuit boards of this size, the heat expansion from the soldering process can cause severe warping of the circuit boards. While I did place a board stiffener on the new printed circuit board to prevent warping during normal handling, it was not strong enough to prevent warping resulting from the soldering process. All four boards were quite visibly warped. Probably because of the warping, one of the four new boards never worked. It might have been possible to fix the bad board with a lot of work, but I had a thesis to defend and lots of packing to do before I left for IBM. Therefore, the maximum number of boards that we could really have for the match was four.

在我最初的积极回应之后,谢尔比回来了有关拟议比赛的更多详细信息。NYNEX(纽约地区的电话公司)将赞助该活动。我向谢尔比提到,我们中的一些人将转向 IBM,但 NYNEX 显然对 IBM 连接没有任何问题。这场比赛的时间很短,只持续一天。比赛将进行两场比赛,每场比赛的比赛时间为 90 分钟。每场比赛最多持续三个小时。加里·卡斯帕罗夫(Garry Kasparov)这场比赛的一日出场费(不包括费用)为 10,000 美元,据说这是大幅折扣。他认为这场比赛更像是一次实验,而不是一场真正的比赛,并且他愿意给予折扣。无论如何,一天工作一万美元仍然是一笔可观的数字。(1997年,我受IBM亚太公司委托,进行为期近一个月的巡回演讲和演讲,途经七个国家。包括机票和酒店住宿在内的总费用略高于10,000美元。现在我相信加里一日出场费约为 30,000 美元,足够我环游世界三个月的商务旅行了。)

After my initial positive response, Shelby came back with more details for the proposed match. NYNEX, the Telephone Company for the New York area, would sponsor it. I mentioned to Shelby that some of us would be moving to IBM, but NYNEX apparently had no problem with the IBM connection. The match would be a short one, lasting only one day. Two games would be played, and the games would be played with 90 minutes to each player for the entire game. Each game would last at most three hours. Garry Kasparov’s one-day appearance fee, not including expenses, for the match would be $10,000, which was supposedly heavily discounted. He viewed the match more as an experiment than as a real match and he was willing to give a discount. Anyway, $10,000 for one day of work is still a tidy sum. (In 1997, I was asked by IBM Asia Pacific to give a lecture and presentation tour for nearly a month, stopping in seven countries. The total expenses, including the plane fare and hotel accommodation, came to slightly more than $10,000. Nowadays I believe that Garry’s fee for a one-day appearance is something like $30,000, enough for me to take a three-month business trip around the world.)

Murray 在卡内基梅隆大学担任助理研究员后于 1989 年 9 月加入 IBM。一个月后,我在论文答辩之后加入了 IBM 2。托马斯距离完成论文还有很长的路要走,因此留在了卡内基梅隆大学。在我们离开大学之前,IBM 和卡内基梅隆大学达成协议,我们可以将三块工作中的 Deep Thought 印刷电路板带到 IBM。由于托马斯留下来,我只带了一块木板。另外两个新板和原来的 Deep Thought 板仍留在卡内基梅隆大学,以便托马斯可以使用所有三个板完成并行代码。这三块板将放置在三个工作站内,每个工作站内放置一块板。然后,工作站将通过卡内基梅隆大学的部门局域网进行通信,并协商如何并行运行搜索。我们在 Deep Thought 与 Kasparov 比赛中使用的就是这种配置。它的运行速度是普通深度思考的两到三倍。它不会与加里竞争,但至少我们可以希望有一个好的表现。

Murray joined IBM in September 1989 after finishing his stint as a research associate at Carnegie Mellon. I joined IBM a month later, after my thesis defense 2. Thomas was still quite a way from completing his thesis and stayed behind at Carnegie Mellon. Prior to our leaving the university, IBM and Carnegie Mellon reached an agreement that we could take the three working Deep Thought printed circuit boards with us to IBM. Since Thomas was staying behind, I brought only one board with me. The other two new boards and the original Deep Thought board remained at Carnegie Mellon so that Thomas could complete the parallel code using all three boards. The three boards would be placed inside three workstations, with one board in each. The workstations would then communicate over the departmental local area network at Carnegie Mellon and negotiate how the search would be run in parallel. This configuration was what we used at the Deep Thought vs. Kasparov match. It ran at two to three times the speed of the regular Deep Thought. It would not be competitive with Garry, but at least we could hope for a good showing.

但卡尔波夫和我不是

But Not Karpov and Me

我第一次从我在卡内基梅隆大学的同事迈克·布朗那里听到加里·卡斯帕罗夫(Garry Kasparov)这个名字,大概是在 1983 年底,当时正值国际象棋世界锦标赛的中期。当时我不知道国际象棋界的大牌是谁。迈克必须向我解释阿纳托利·卡尔波夫是世界冠军。维克托·科尔奇诺伊(Victor Korchnoi)是可能的挑战者之一,据一些人称,他在非常不利的条件下输给了阿纳托利(Anatoly)两场冠军赛。维克多叛逃,他在苏联的家人因此受到一些迫害。有传言说,如果维克多赢了其中一场比赛,他的家人就会受到折磨。迈克认为维克多将是最有可能的挑战者,但他不相信自己能战胜阿纳托利。迈克随后提到了加里·卡斯帕罗夫(Garry Kasparov)这个名字,形容他是一位才华横溢的攻击手。迈克认为加里会成为比阿纳托利更令人兴奋的世界冠军,而阿纳托利通常被认为有点无聊。但是,迈克再次认为加里太年轻,太缺乏经验,也无法击败卡尔波夫。

I first heard the name Garry Kasparov from Mike Browne, my office mate at Carnegie Mellon, probably in late 1983, in the middle of a World Chess Championship cycle. At the time I had no idea who the big names in chess were. Mike had to explain to me that Anatoly Karpov was the World Champion. Among the possible challengers was Victor Korchnoi, who had lost two championship matches to Anatoly under, according to some, highly unfavorable conditions. Victor had defected and his family in the USSR suffered some persecution as a result. There was a rumor that Victor’s family would be tortured if he won either of the matches. Mike thought Victor would be the most likely challenger, but he did not believe that he could win against Anatoly. Mike then mentioned the name Garry Kasparov, describing him as a flamboyant attacker with lots of talent. Mike thought that Garry would be a more exciting world champion than Anatoly, who was generally considered a little bit boring. But, again, Mike did not think Garry, being too young and too inexperienced, would be able to beat Karpov either.

1984年,加里从候选人三场比赛中脱颖而出,成为阿纳托利的挑战者。1984年9月9日,他们之间的第一场世界冠军赛开始了。谁先获得六场胜利,谁将被宣布为这场比赛的获胜者。阿纳托利在前九场比赛中很快取得了四场胜利,而加里却未能赢得一场比赛。对于加里来说,情况看起来很黯淡。然后他埋头苦干,一场又一场地平局。几周变成了几个月,记者们开始离开。但比赛继续进行,阿纳托利在第27局再次取得胜利。加里赢得了第32局,然后是第47局和第48局。

In 1984, Garry emerged from the Candidate matches3 as the challenger to Anatoly. On 9 September 1984, the first World Championship match between them began. Whoever scored six wins first would be declared the winner of the match. Anatoly quickly scored four wins in the first nine games, while Garry was not able to win a single game. It looked bleak for Garry. Then he dug in, and drew game after game. Weeks grew into months, and the reporters started to leave. But the match went on, and Anatoly scored another win in game 27. Garry won the 32nd game and then the 47th and the 48th games.

1985年2月15日,国际棋联主席弗洛伦西奥·坎波马内斯宣布“比赛在没有最终结果的情况下终止”。加里抱怨坎波马内斯与阿纳托利是联盟,并暂停了比赛,让阿纳托利得到急需的休息。一场新的 24 场比赛将于 1985 年 9 月 2 日开始,初始比分是 0-0,平局各算半分。如果 24 场比赛后打平,阿纳托利将保住冠军头衔。我很忙正在研究着棋生成器芯片,没有注意比赛。这场比赛加里以13比11的比分获胜,成为世界冠军。

On 15 February 1985, the FIDE (the international chess federation) president Florencio Campomanes announced “the termination of the match without a conclusive result”. Garry complained that Campomanes was in league with Anatoly and was stopping the match to give Anatoly much needed rest. A new 24-game match, with the initial score of 0–0 and with draws counted as half a point each, would begin on 2 September 1985. In the case of a tie after 24 games, Anatoly would retain his title. I was busy working on the move generator chip and was not paying attention to the match. In this match Garry prevailed by the score of 13 points to 11 points, and became the World Champion.

到1989年10月,加里已经两次对阵阿纳托利,成功卫冕。他即将正式成为第一位、也是迄今为止唯一一位国际棋联评分达到 2800 或以上的国际象棋棋手。此前 2780 的收视率记录保持者正是传奇人物鲍比·费舍尔 (Bobby Fischer)。

By October 1989, Garry had successfully defended his title twice against Anatoly. He was officially about to become the first and, to the date of writing, the only chess player to have a FIDE rating of 2800 or above. None other than the legendary Bobby Fischer held the previous rating record of 2780.

在 Deep Thought 和 Garry 比赛的前几天,我们接到了哥伦比亚广播公司 (CBS) 的电话。《CBS This Morning》想要采访 Garry 和 Deep Thought 团队。IBM 研究部通信部门负责国际象棋项目公共关系的 Jerry Present 与 CBS 进行了详细安排。默里和我在纽约市的一家酒店过夜,以便我们能够准时到达工作室。

A few days before the scheduled match between Deep Thought and Garry, we got a call from CBS. “CBS This Morning” wanted to do an interview with Garry and the Deep Thought team. Jerry Present, who handled the public relations of the chess project for the Communications department at IBM Research, made the detailed arrangements with CBS. Murray and I stayed overnight at a hotel in New York City so that we could get to the studio on time.

我们起得很早,早上5点左右就到达了演播室,距离演出开始还有两个多小时。CBS 希望我们从演播室远程连接 Deep Thought,并在该片段中演奏一些动作。设置完 IBM PC 并测试与 IBM Research 的电话连接后,我们被带到等候室吃早餐。加里还没有到。房间里的几台显示器正在播放哥伦比亚广播公司、其他网络和当地电视台的直播节目。哥伦比亚广播公司的一位工作人员留下来给我们提供了一些内部评论。这种情况持续了一段时间。然后我们听到了脚步声。门口出现了一张熟悉的面孔。是加里。

We woke early and arrived at the studio at about 5 o’clock in the morning, more than two hours before the show was to start. CBS wanted us to hook up Deep Thought remotely from the studio and play a few moves during the segment. After setting up an IBM PC and testing the phone connection to IBM Research, we were led to a waiting room to have some breakfast. Garry had not arrived yet. Several monitors in the room were showing the on-air shows from CBS, other networks and the local TV stations. One of the CBS people stuck around and gave us some insider comments. This went on for a while. Then we heard the sound of footsteps. A familiar face appeared at the door. It was Garry.

从我看到的照片来看,Garry 看起来是一个精力充沛、精力充沛的人。亲眼见到他证实了这个评估。然而,照片无法真正表达他的存在所带来的强度。漆黑的眸子后面,闪烁着熊熊的火焰。加里曾经将下棋称为“控制混乱”。那双眼睛背后是否正在发生秩序与混乱之间的战斗?

From the photos that I had seen, Garry appeared to be an intense person with a very high energy level. Seeing him in person confirmed the assessment. Photographs, however, could not really convey the intensity that came from his presence. Behind his dark eyes, there was a smoldering fire. Garry once referred to playing chess as “controlling chaos”. Was there a fight going on between order and chaos behind those eyes?

我们自我介绍并聊了一会儿。哥伦比亚广播公司的一名工作人员带领加里去化妆。我们很快就跟了上去。化妆结束后,我们的话题自然而然地转向了国际象棋。加里即将取代鲍比·费舍尔成为有史以来评分最高的国际象棋棋手,我们讨论了比较两个不同时期评分的有效性。卡斯帕罗夫拒绝直接评论两人之间的相对国际象棋实力他自己和鲍比。我们没有费心去谈论加里和深思之间的比赛可能的结果——毫无疑问他会轻松获胜。但是 Garry 认为他对我们将要在 IBM 制造的机器的机会是什么?他做了一个有趣的观察。

We introduced ourselves and chatted a little bit. A CBS person led Garry away to his makeup session. We followed shortly. After the makeup sessions, our conversation naturally turned to chess. Garry was about to overtake Bobby Fischer as the highest rated chess player ever, and we talked about the validity of comparing ratings from two different time periods. Kasparov declined to comment directly about the relative chess playing strength between himself and Bobby. We did not bother to talk about the likely outcome of the match between Garry and Deep Thought—there was never any doubt that he would win it easily. But what did Garry think his chance was against the machine that we were going to build at IBM? He made an interesting observation.

当团队加入 IBM 时,我们计划构建一台每秒至少可以搜索 1 亿个位置的国际象棋机器,最好是每秒搜索 10 亿个位置。加里认为每秒搜索 1 亿个位置的机器将是一个有趣的对手,并且可能会击败大多数大师。但即使对于每秒 10 亿个位置的机器,加里表示,“它可能能够击败其他大师,但无法击败阿纳托利和我。” 然后他解释说,在他看来,他和阿纳托利·卡尔波夫是世界上唯一真正的职业棋手,也是唯一能够为大赛做好适当准备的人。阿纳托利的评分为2730,紧随其后的是蒂曼、格尔凡德和伊万丘克。他们的评分约为 2680 分,略低于阿纳托利,但还没有达到全班(200 分)。如果只看评分标准,没有理由相信这款新机器会落在阿纳托利和其他顶级通用汽车之间的狭窄范围内。那么加里的真正意思是什么?

When the team joined IBM, we were planning to build a chess machine that could search at least 100 million positions per second, preferably a billion positions per second. Garry thought that a machine searching 100 million positions per second would be an interesting opponent, and probably would beat the majority of the Grandmasters. But even for a one billion positions per second machine, Garry stated, “It might be able to beat the rest of the Grandmasters, but not Anatoly and me.” He then explained that, in his opinion, he and Anatoly Karpov were the only real professional chess players in the world, and that they were the only ones able to do proper preparations for the big match. Anatoly was rated 2730, and the players right behind him were Timman, Gelfand and Ivanchuk. They were rated at about 2680, slightly behind Anatoly, but not by a full class (200 rating points). If one just looked at the rating scale, there was no reason to believe that the new machine would fall in the narrow range between Anatoly and the other top GMs. So what did Garry really mean?

有两种看似合理的理论。

There are two plausible theories.

一是加里潜意识里知道这台机器将是一个严峻的挑战,但他拒绝向不可避免的事情低头,他将阿纳托利放在机器和他自己之间作为缓冲区。这是默里第一次读到弗里德尔信中加里“不是卡尔波夫和我”的说法时的理论。纵观计算机国际象棋的历史,国际象棋棋手也曾做出过类似的断言,但 Deep Thought 打破了所谓的障碍,一直到“强大的大师”。有趣的是,在1997年深蓝与Garry的比赛中,几位大师声称深蓝的世界排名在X左右。X的数字恰好略低于提出要求的特定大师的排名。一些提出这一主张的大师可能会忙于古老的深思。

One was that Garry knew subconsciously the machine would be a serious challenge but, refusing to bow to the inevitable, he placed Anatoly between the machine and himself as a buffer. This was Murray’s theory when he first read about Garry’s “Not Karpov and me” statement in Friedel’s letter. Throughout the history of computer chess, chess players had made similar assertions but Deep Thought smashed the supposed barriers all the way up to “strong Grandmasters”. It was interesting to note that, during the 1997 match between Deep Blue and Garry, several Grandmasters claimed that Deep Blue was playing at around number X in the world. The number X happened to be slightly below the ranking of the specific Grandmaster making the claim. Some of the Grandmasters making the claims probably would have had their hands full with the old Deep Thought.

另一个看似合理的理论是,加里确实相信良好的准备会对比赛表现产生巨大的影响。其实我们也相信这一点。事实上,“终极国际象棋机器”的设计速度是经过指定的,以便机器能够比毫无准备的加里或世界冠军强约 200 分。我们有充分的理由去超越。准备充分的 IM David Levy 能够以 4 比 0 的比分击败了 Cray Blitz,而 Cray Blitz 的国际象棋实力应该与 David 相当。因此,在面对缓慢变化的电脑对手时,良好的准备确实可以发挥巨大作用。但为什么是“卡尔波夫和我”呢?

The other plausible theory was that Garry really believed good preparation would make a huge difference in match performance. Actually, we believed this as well. In fact, the design speed for the “ultimate chess machine” was specified so that the machine would play at about 200 points stronger than an unprepared Garry, or whoever the World Champion might be. We had good reasons to overshoot. A well-prepared IM David Levy was able to slaughter Cray Blitz, which was supposed to be comparable to David in chess strength, by 4 to 0. So it was true that good preparation could make a huge difference against a slowly changing computer opponent. But why “Karpov and me”?

其他大师无法做出类似的准备并没有根本原因,但加里和阿纳托利确实比其他大师拥有更多的资源。大师乔尔·本杰明(Joel Benjamin)在两场深蓝对阵加里的比赛中与我们一起工作,他观察到加里和阿纳托利有很多“奴隶”。许多特级大师都希望加入冠军阵营,这样他们就可以过上体面的生活,或者在某些情况下,了解如何成为世界冠军。在苏联,一个大师要想过上好日子,除非他是位高权重的人,否则是很难的,而苏联的解体和随后失去的国家支持使情况变得更加严峻。加里和阿纳托利都是千万富翁,都吸引了前苏联大师。西方大师,比如来自荷兰的简·蒂曼,要获得同样质量的帮助要困难得多。此外,加里和阿纳托利都成功卫冕。谁敢说他们没有一些顶级准备的秘方?

There was no fundamental reason why the other Grandmasters could not make comparable preparation but Garry and Anatoly did have more resources at their disposal than other Grandmasters. Grandmaster Joel Benjamin, who worked with us during the two Deep Blue matches against Garry, observed that Garry and Anatoly had lots of “slaves”. Numerous Grandmasters wanted to be in the Champion’s camp so that they could eke out a decent living or, in some cases, learn what it would take to be a World Champion. It was very difficult for a Grandmaster in the USSR to make a good living unless he was right at the top, and the collapse of the USSR and the subsequent loss of state support made the situation even more acute. Both Garry and Anatoly, each being a multimillionaire, became magnets to the ex-USSR Grandmasters. Western Grandmasters, like Jan Timman from the Netherlands, would have a much tougher time getting the same quality help. Furthermore, Garry and Anatoly had both successfully defended their titles. Who is to say that they don’t have some secret formulae for top level preparation?

电视直播环节进展顺利。加里和深思下了一些象征性的棋。毫不奇怪,采访的其余部分都集中在加里身上。毕竟,默里和我都只是刚从研究生院毕业的微不足道的学生。从他的话中,我的印象是,他认为自己是在捍卫人类免受机器的侵害。当我看着我的团队成员时,我看到了人的面孔。我们只是从一个非传统的方向来解决下棋的问题。也许加里需要为即将到来的比赛打气。嗯,这对我来说很好。我们希望他发挥出最佳状态——好吧,尽管也许不是为了即将到来的 Deep Thought 比赛,因为我们一开始就处于劣势。

The live TV segment went smoothly. Garry and Deep Thought played a few token moves. The rest of the interview, not surprisingly, was centered on Garry. Murray and I were, after all, just insignificant students fresh out of graduate school. My impression, from what he said, was that he saw himself as defending the human race against machines. When I looked at my fellow team members, I saw human faces. We were simply approaching the problem of playing chess from, shall we say, a non-traditional direction. Maybe Garry needed to pump himself up for the matches to come. Well, that was fine with me. We wanted him to play at his best—okay, though perhaps not for the coming Deep Thought match where we were overmatched to begin with.

谁想赢?

Who Wants to Win?

毫无疑问,加里想以2-0的完美比分赢得这场与深思的小型表演赛。我们,在另一方面,我们只关心不要让自己难堪。0-2 输球对我们来说没什么大不了的;说实话,深思两败甚至可能是最理想的结果。我们当然不希望我们的新雇主抱有不切实际的期望,认为我们已经准备好真正应对世界冠军了。然而,谢尔比·莱曼不确定他想看到什么结果。加里 2-0 获胜固然不错,但 Deep Thought 的任何半分平局都会成为头条新闻。

There was no doubt that Garry wanted to win the small exhibition match against Deep Thought with a perfect 2–0 score. We, on the other hand, were only concerned about not embarrassing ourselves. Losing by 0–2 would not be a big deal for us; to be honest, two losses by Deep Thought might even be the most desirable outcome. We certainly did not want our new employer to have the unrealistic expectation that we were ready to tackle the World Champion for real. Shelby Lyman, however, was not sure what outcome he wanted to see. A 2–0 win by Garry would be fine, but any half-point draw by Deep Thought would create huge headlines.

1989 年 10 月 21 日星期六,也就是我们第一次与加里进行全面交锋的前一天,我们参观了比赛场地并检查了设施。这场比赛将在纽约艺术学院举行,对于电脑国际象棋比赛来说,这似乎是一个奇怪的地方。当我们走进游戏室时,映入眼帘的是一幅不寻常的景象。到处都是绘画和雕塑。一个角落里有一只巨大的手,大约有双人沙发那么大,可能是石膏制成的。难道这就是命运之手吗?或者它是否象征着“深思”,这个我们亲手创造的工具,在与加里的斗争中?

On Saturday 21 October 1989, the day before our first over-the-board encounter with Garry, we visited the match site and checked out the facilities. The match was to take place at the New York Academy of Art, which seemed a strange place for a computer chess match. When we walked into the playing room, we were greeted by an unusual sight. There were paintings and sculptures everywhere. In one corner was a giant hand, roughly the size of a love seat, possibly made from gypsum. Was it supposed to be the hand of fate? Or did it symbolize Deep Thought, the tool we had created with our hands, in its struggle with Garry?

比赛场地相当局促,尤其是加了所有观众椅后,给人一种亲密的感觉。加里在小舞台后面有一个房间可以休息,舞台后面还有一个小地方可以让他一边走一边思考。尽管当天没有安排新闻发布会,但仍有一些记者在周围闲逛。我们设置了一台 IBM PC,连接了电话线,并测试了与匹兹堡的连接。其中一名记者想要采访。原来他来自《体育画报》。我们已经不在堪萨斯了。

The playing room was quite cramped, especially after all the chairs for the spectators were added, giving an intimate feeling. Garry had a room behind the small raised stage to rest in and there was also a small area behind the stage that he could stroll around while he was thinking. A few reporters were floating around even though no press conference was scheduled for that day. We set up an IBM PC, hooked up the phone line, and tested the connection to Pittsburgh. One of the reporters wanted an interview. It turned out he was from Sports Illustrated. We were not in Kansas any more.

接受《体育画报》采访后我们上楼找到谢尔比·莱曼。观众可以听到现场解说的解说室在二楼,但里面空无一人。当我们找到他时,谢尔比非常激动。加里似乎拒绝与一名记者交谈,因为这名记者似乎做了一些让他不高兴的事情。谢尔比和其他几个人试图让加里改变主意。我很好奇加里的母亲克拉拉·卡斯帕洛娃是否在附近,但谢尔比表示克拉拉没有来纽约。因此,正如我所怀疑的那样,加里并不认为这场比赛是一场重大赛事。克拉拉的丈夫在加里还是个小男孩的时候就去世了,他和他的母亲关系非常密切。在他的自传中,加里说:“最重要的是我可以和她(他的母亲)交谈,就像我不能和其他人一样。” 如果这是一件大事,克拉拉就会出现为她的儿子加油。然而,正如我们很快就会发现的那样,加里仍然为比赛做了充分的准备。

After the interview with Sports Illustrated, we went upstairs to find Shelby Lyman. The commentary room, where the spectators could hear the live commentary, was on the second floor, but it was empty. When we found him, Shelby was quite agitated. It appeared that Garry had refused to talk to one of the reporters who appeared to have done something to upset him. Shelby and a few other people were trying to make Garry change his mind. I was curious as to whether Garry’s mother, Clara Kasparova, was around but Shelby indicated that Clara had not come to New York. So, as I suspected, Garry did not consider the match a major event. Clara’s husband had died when Garry was still a little boy and he and his mother had a very close relationship. In his autobiography, Garry said, “The most important thing is that I can talk to her [his mother] as I can to nobody else.” If it had been a major event, Clara would have shown up to cheer her son on. Garry, however, still did a thorough preparation for the match, as we would soon find out.

当我们周日到达比赛场地时,我们看到了一些在国际象棋比赛中通常看不到的东西。排队买票的人绕了整个街区。大约有三十到五十名记者在加里周围转来转去。一些人注意到我们并开始问我们问题。过了一会儿,记者们就座,准备参加赛前新闻发布会。默里和我进行了简短的讨论,并一致认为,如果 Deep Thought 对该头寸的评估跌破 –700,即大约下跌 5.5 个棋子4,那么我们将辞职。讨论结束后,我抬起头,看到加里对我们微笑。他无意中听到了谈话。

When we arrived at the venue on Sunday, we saw something that you don’t normally see for a chess match. The line of people waiting to buy tickets circled an entire city block. Around thirty to fifty reporters were milling around Garry. A few noticed us and started asking us questions. After a while, the reporters got seated and ready for the pre-match press conference. Murray and I had a brief discussion and agreed that if Deep Thought’s assessment of the position dropped below –700, which was roughly 5.5 pawns down 4, then we would resign. After the discussion I looked up and saw Garry smiling broadly at us; he had overheard the conversation.

除了记者之外,还有几位电脑象棋比赛中常见的嫌疑人也在场。坐在前排的是 ACM 电脑象棋锦标赛的常年组织者 Monty Newborn。我们的前竞争对手 Hans Berliner 博士也是如此,他显然不想错过这一历史性的场面,尽管他可能希望 Hitech 成为计算机的竞争对手。国际象棋节目 Phoenix 的作者 Jonathan “Almost Winning” Schaeffer 也在场。(Deep Thought 曾与 Phoenix 交手过 3 次,三场均获胜。但根据 Jonathan 的说法,Phoenix 总是“几乎获胜”。)ACM 和世界计算机象棋锦标赛的赛事总监 Mike Valvo 也出席了。

Besides the reporters, several of the usual suspects from computer chess events were also present. Monty Newborn, the perennial organizer of the ACM Computer Chess Championship, was in the front row. So was our former rival, Dr Hans Berliner, apparently not wanting to miss the historic spectacle even though he probably wished that Hitech was the computer competitor. Jonathan “Almost Winning” Schaeffer, author of the chess program Phoenix, was also present. (Deep Thought had played Phoenix three times, winning all three. But according to Jonathan, Phoenix was always “almost winning”.) Mike Valvo, the tournament director for the ACM and the World Computer Chess Championship, also showed up.

第一局Deep Thought下白,我是操手。穆雷负责评论室,通过电脑转播动作。反过来,PC 则跟踪匹兹堡计算机上的游戏日志。

Deep Thought got White in the first game, and I was the operator. Murray manned the commentary room, relaying the moves from a PC. The PC, in turn, was following the game log on the computer in Pittsburgh.

Deep Thought 的比赛开局良好。然后我注意到发生了一些奇怪的事情。连续数步,深思想要攻克王,但随后又转向毫无意义的着法。最终,深思决定真正建立城堡,但由于奇怪的举动,加里已经占据了优势。

The game started well for Deep Thought. Then I noticed that something strange was happening. For several moves in a row, Deep Thought wanted to castle its king, but then switched to moves that did not make any sense. Eventually, Deep Thought decided to castle for real, but Garry already had an edge as the result of the strange moves.

三周后,Deep Thought 在 ACM 锦标赛中对阵 Phoenix。同样的行为再次发生,深思在对阵菲尼克斯的比赛中客观上输了。(所以乔纳森这次“几乎获胜”的说法是正确的。)对我们来说幸运的是,Phoenix只是“几乎获胜”,而Deep Thought在残局中经过长时间的挣扎后赢得了比赛。托马斯在凤凰城的叫醒电话后发现了问题。在急于让三计算机版本在加里·卡斯帕罗夫(Garry Kasparov)比赛中及时运行时,引入了一个微妙的错误。如果在搜索的软件部分5中立即进行易位移动,则Deep Thought对该位置的评估不会获得易位奖励。但如果易位移动被延迟,并发挥到搜索的硬件部分,那么深思将获得奖金。因此,Deep Thought 会采取极端的措施来延迟易位,直到它进入搜索的硬件部分。在凤凰城的比赛中,深思最终没有到达城堡。乔纳森在比赛的大部分时间里以我为代价度过了一段非常愉快的时光。用他自己的话说,我“坐立不安、愁眉苦脸、忧心忡忡” 6。但当形势对 Deep Thought 有利后,我“头晕目眩、面带微笑、精力充沛”。乔纳森对这款游戏非常着迷,很快就放弃了电脑国际象棋7。在凤凰城比赛之后,我们永久停止使用三电脑版​​本的 Deep Thought。

Three weeks later, Deep Thought played Phoenix in the ACM Championship. The same behavior happened again, and Deep Thought was objectively lost against Phoenix. (So Jonathan was correct in his “almost winning” assertion this time.) Luckily for us, Phoenix was only “almost winning”, and Deep Thought won the game after a long struggle in the endgame. Thomas found the problem after the wakeup call from Phoenix. In the rush to get the three-computer version working in time for the Garry Kasparov match, a subtle bug had been introduced. If the castling move was made immediately in the software portion 5 of the search, then no castling bonus was awarded to Deep Thought’s evaluation of the position. But if the castling move was delayed, and played into the hardware part of search, then Deep Thought would get the bonus. So Deep Thought would go to extreme contortions to delay castling until it was in the hardware portion of the search. In the Phoenix game, Deep Thought never did get to castle as a result. Jonathan had a very pleasant time for most of the game at my expense. In his own words, I was “fidgeting, scowling, concerned”6. But after the position turned in Deep Thought’s favor, I was “giddy, smiling and exuberant”. Jonathan took the game quite hard and soon gave up on computer chess 7. We stopped using the three-computer version of Deep Thought permanently after the Phoenix game.

无论如何,在与加里的第一场比赛中,深思的棋子在荒唐的扭曲之后严重错位。加里慢慢地增加了他的位置优势,在52步之后,当Deep Thought的评价跌破-700时,我退出了Deep Thought。比赛在这之前就已经输掉了。

Anyway, in the first game against Garry, Deep Thought’s pieces were seriously misplaced after the ridiculous contortions. Garry slowly increased his positional advantage and, after 52 moves, when Deep Thought’s evaluation dropped below −700, I resigned for Deep Thought. The game was lost way before that.

当第一场比赛正在进行时,PBS 的摄制组电视节目“Nova”正在拍摄加里。第一局比赛结束后,Nova 制作人提到 Garry 盯着我看,“就好像你是敌人一样”。我一心一意不犯操作错误,以至于自己没有注意到——有时,无知就是幸福。

While the first game was going on, a filming crew for the PBS TV show “Nova” was taking shots of Garry. After the first game, the Nova producer mentioned that Garry was staring at me “as if you were the enemy”. I had been so intent on not making operator errors that I had not noticed—sometimes, ignorance is bliss.

加里、穆雷和我在第二场比赛中交换了位置。穆雷面对加里,而我将动作转述给评论员。加里有怀特。

Garry, Murray and I traded places for the second game. Murray got to face Garry, while I relayed the moves to the commentators. Garry had White.

赛前,加里要求我们将 Deep Thought 的所有公开比赛发送给他。我们很高兴地答应了。我们并没有真正认真地对待这场比赛。如果这是一场真正的比赛,我们所做的不仅仅是让加里参加公开比赛。如果需要的话,我们会进行培训,准备新的空缺职位,并彻底改变计划甚至硬件。人类国际象棋棋手在参加世界锦标赛时也会进行同样类型的秘密训练。通常,两个阵营都会在重大比赛前六个月内避免正式公开露面。显然,两个阵营都不会主动给对手准备任何惊喜。当加里参加世界锦标赛时,他与阿纳托利·卡尔波夫、奈杰尔·肖特(英国)和维什瓦纳坦·阿南德(印度)的比赛就是这种情况。然而,在这场比赛中,“深思”对加里来说是一本开放的书,我们在第二场比赛中为这种开放性,或者更确切地说,缺乏秘密训练付出了高昂的代价。

Before the match, Garry had requested that we send him all of Deep Thought’s public games. We gladly obliged. We weren’t really taking the match seriously. If it had been a genuine match, we would have done more than just give Garry the public games. We would have gone into training, prepared new openings, and drastically changed the program or even the hardware, if need be. Human chess players do the same type of secret training when they go into World Championship matches. Usually, both camps shy away from serious public appearances for, say, six months before the big match. Obviously, neither camp would volunteer any surprises that they had prepared for their opponents. This was the case when Garry played for the World Championship in his matches against Anatoly Karpov, Nigel Short (Great Britain) and Vishwanathan Anand (India). In this match, however, Deep Thought was an open book to Garry, and we paid dearly for this openness, or rather, lack of secret training, in the second game.

加里使用了一系列的动作,打破了黑方深思的空当。至少可以说,开场后的局面对于 Deep Thought 来说是很困难的。柏林博士非常喜欢怀特的位置,因此他准备并扩展了加里的路线,供高科技公司用来对抗深思。加里·卡斯帕罗夫 (Garry Kasparov) 比赛三周后,Hitech 在同一开场对阵 Deep Thought 的比赛中,白队出局。Deep Thought在开局后就惨遭失败,但Hitech的实力不足以终结Deep Thought。Hitech“几乎获胜”。但一经出版,Hitech 就立即误入歧途。深思力挽狂澜,赢得了比赛。我们对阵加里就没那么幸运了。在与加里的两场比赛中都存在的易车错误也没有帮助。深思一开始就意识到自己遇到了麻烦,而且从那里开始就直线走下坡路了。Deep Thought 从未想过自己在游戏中处于领先地位,即使它的材料一度领先。然而,现场评论声称 Deep Thought 相信它处于领先地位。一本描述这款游戏的书甚至更深入地描述了这款游戏据说Deep Thought的评价是这样那样的超前。基于Deep Thought在整场比赛中对其位置的高度负面评价,我相信Garry只是简单地破坏了Deep Thought的开局。穆雷在第37步为Deep Thought认输,但在灾难性的开局之后,比赛的结果就不再悬而未决了。

Garry used a move sequence that busted Deep Thought’s opening for the Black side. The position right after the opening was difficult for Deep Thought, to say the least. Dr Berliner liked the position for White so much that he prepared and extended Garry’s line for use by Hitech against Deep Thought. Three weeks after the Garry Kasparov game, Hitech played the White side of the same opening against Deep Thought. Deep Thought was lost right after the opening, but Hitech was not strong enough to finish off Deep Thought. Hitech was “almost winning”. But once it was out of its opening book, Hitech immediately went astray. Deep Thought turned the tide and won the game. We did not have such luck against Garry. The castling bug, which existed in both games against Garry, did not help either. Deep Thought realized that it was in trouble right out of the opening, and it was straight downhill from there. Deep Thought never thought it was ahead in the game even when it was up in material at one time. The on-site commentary, however, was asserting that Deep Thought believed it was ahead. One book that described the game even went as far as saying Deep Thought’s evaluation was such and such ahead for it. Based on Deep Thought’s highly negative evaluation of its positions during the entire game, I believe that Garry simply busted Deep Thought’s opening. Murray resigned for Deep Thought on move 37, but the result of the game was never in doubt after the disastrous opening.

加里第二场比赛开场准备的彻底程度让我们大开眼界。Deep Thought从来没有真正玩过这个游戏。比赛输给了加里的主场8。然而,第二场比赛的轻松获胜可能会影响加里对他在与深蓝的两场比赛中的预期的判断。

The thoroughness of Garry’s opening preparation for the second game was an eye opener for us. Deep Thought never got to really play the game. It lost the game to Garry’s home preparation 8. The easy win in the second game, however, might later have clouded Garry’s judgment of what he could expect in the two matches he played against Deep Blue.

卡尔波夫先生将有白色

Mr Karpov Will Have White

加里·卡斯帕罗夫(Garry Kasparov)在与“深思”(Deep Thought)的比赛中获胜,为他带来了迄今为止美国最大的媒体报道。赛后,加里的私人经理安德鲁·佩吉9 (Andrew Page 9 ) 对加里受到的热烈欢迎感到非常高兴。加里冒着风险10同意了出场费的大幅折扣,但轻描淡写地说,他所获得的宣传完全值得这个折扣。在某些方面,加里从 Deep Thought 比赛中获得的报道甚至比他的世界锦标赛比赛还要多。

Garry Kasparov’s victory in the match against Deep Thought gave him by far the largest media coverage ever in the United States. After the match, Andrew Page9, Garry’s personal manager, was extremely happy about how well Garry was received. Garry took a risk10 by agreeing to the heavy discounted appearance fee, but it would be an understatement to say that the publicity he received was well worth the discount. In certain aspects, Garry was getting even more coverage from the Deep Thought match than his World Championship matches.

加里的媒体财富并没有被其他国际象棋棋手忽视,“深思”很快就会在表演赛中与其他三名国际象棋棋手较量。三人的最后一场比赛是对阵德国大师赫尔穆特·普弗莱格。这场比赛的结果是两场平局。另外两场比赛的结果更为精彩。

Garry’s media bonanza did not go unnoticed by the other chess players, and Deep Thought would soon take on three other chess players in exhibition matches. The last match of the three was against German Grandmaster Helmut Pfleger. The match resulted in two draws. The other two matches had more exciting results.

其中第一场比赛于 1989 年 12 月举行,距加里·卡斯帕罗夫 (Garry Kasparov) 比赛大约两个月后。这场对阵国际大师大卫·利维的比赛的安排实际上是在与卡斯帕罗夫的比赛之前就开始了。大卫因嘲笑计算机国际象棋研究人员而闻名。1968年,他与几位计算机科学家打赌,未来10年内没有计算机会在比赛中击败他。他收集了这是 1978 年西北大学的国际象棋程序 Chess 4.5 输给他时的赌注。Omni杂志随后设立了 4000 美元的 Omni 奖,David 又增加了 1000 美元,奖励第一台在比赛中击败他的计算机。1984年,Cray Blitz团队试图夺得全能奖,但被David以4-0的满分惨败。

The first of these matches took place in December 1989, about two months after the Garry Kasparov match. The arrangements for this match, against International Master David Levy, were actually started before the Garry Kasparov match. David was best known for his taunting of computer chess researchers. In 1968, he had placed a bet with several computer scientists that no computers would beat him in a match within the next 10 years. He collected the wager when Chess 4.5, a chess program from the Northwestern University, lost to him in 1978. Omni magazine then set up a $4000 Omni Prize, and David added another $1000 to the pile for the first computer to defeat him in a match. The Cray Blitz team tried to claim the Omni Prize in 1984, but was soundly defeated by David with the perfect score of 4–0.

当大卫玩克雷闪电战时,他采用了他所谓的“什么都不做,但要做好”的策略。他刻意避免给克雷闪电战目标,同时保持棋盘位置封闭,这样立即采取战术的机会就很小。这个策略在对抗克雷闪电战时非常有效。就此而言,它在对抗许多商业国际象棋程序时表现得相当好,即使是在像我这样的弱国际象棋棋手的手中。对阵大卫时,克雷·闪电战会得过且过,直到他制定出一些不可阻挡的长期战术,然后克雷·闪电战就会崩溃并燃烧。Deep Thought 毫无问题地发现了对抗 Cray Blitz 的长期战术,但考虑到 David 能够轻松地消灭 Cray Blitz,我们在比赛开始前就非常担心。当 Deep Thought 对阵 Garry 时,我们没有得分也是可以接受的。对阵大卫,我们绝对想赢。在 Garry Kasparov 比赛和 David Levy 比赛之间,我们在准备过程中向 Deep Thought 的评估函数添加了 30 个新项。

When David played Cray Blitz, he adopted what he called the “Do nothing, but do it well” strategy. He deliberately avoided giving Cray Blitz targets, while keeping the board position closed so that the chances of immediate tactics were remote. The strategy worked exceedingly well against Cray Blitz. For that matter, it worked quite well against a lot of commercial chess programs, even at the hand of a weak chess player such as myself. Against David, Cray Blitz would muddle along until he set up some unstoppable long-term tactics and then Cray Blitz would crash and burn. Deep Thought had no problems spotting the long-term tactics used against Cray Blitz but, given that David was able to dispatch Cray Blitz with great ease, we were quite concerned before the match took place. When Deep Thought played Garry, it was acceptable that we did not score any point. Against David, we definitely wanted to win. Between the Garry Kasparov match and the David Levy match, we added thirty new terms to Deep Thought’s evaluation function in preparation.

与 David 的比赛是在 Deep Thought 对阵 Phoenix 的比赛几乎遭遇灾难之后进行的,我们又回到了带有两个国际象棋处理器的单台计算机。比赛在伦敦举行。Peter Jansen是操作员,他于 1989 年初参与 Deep Thought 项目11 。理论上,这场比赛可以使用 IBM 的机器进行,但由于只有匹兹堡的机器可供所有 Deep Thought 团队成员使用,因此我们使用卡内基梅隆大学的工作站进行比赛。

The match against David took place after Deep Thought’s near disaster against Phoenix, and we were back to the single computer with two chess processors. The match was played in London. Peter Jansen, who became involved with the Deep Thought project in early 198911, was the operator. In theory, the match could have been played using the machine in IBM, but since only the machines in Pittsburgh were available to all the Deep Thought team members, we used a workstation at Carnegie Mellon for the match.

让我们松了一口气的是,Deep Thought 能够对 David 做他对 Cray Blitz 所做的事情。0-4惨败后,大卫表现得很客气,没有为输球找借口。用他自己的话说,“我并没有感到特别不适;我没有坐得太近加热器;游戏室外没有任何令人不安的噪音;灯光还不错;电脑操作员彼得·詹森没有吸烟。”

Much to our relief, Deep Thought was able to do to David what he had done to Cray Blitz. David was gracious after his 0–4 drubbing, and did not make excuses for losing. In his own words, “I was not feeling particularly unwell; I was not sitting too close to the heater; there was no disturbing noise outside the playing room; the lighting was not poor; the computer operator, Peter Jansen, was not smoking.”

Deep Thought的下一场比赛对手不是别人,正是前国际象棋世界冠军阿纳托利·卡尔波夫。由于加里在 Deep Thought 比赛中获得了如此多的关注,阿纳托利可能也希望获得他的关注。这就是他答应哈佛国际象棋俱乐部邀请下棋的原因吗?我不知道。我确实知道,哈佛国际象棋俱乐部主席丹尼尔·埃德尔曼(Daniel Edelman)邀请 Deep Thought 与阿纳托利对战,是在加里·卡斯帕罗夫(Garry Kasparov)与 Deep Thought 的比赛结束后才发出的。

Deep Thought’s next match opponent was none other than the former World Chess Champion Anatoly Karpov. With Garry getting so much publicity as a result of the Deep Thought match, Anatoly probably wanted his share of the attention. Was this the reason why he agreed to the invitation from the Harvard Chess Club to play Deep Thought? I don’t know. I do know that the invitation from Daniel Edelman, the president of the Harvard Chess Club, for Deep Thought to play Anatoly did not come until after the Garry Kasparov vs. Deep Thought match.

在比赛之前,加里要求提供 Deep Thought 的所有公开比赛的比赛成绩。阿纳托利没有提出同样的要求,我想他一定已经知道了比赛的比分。我大错特错了。

Before his match Garry had requested the game scores of all of Deep Thought’s public games. Anatoly did not make the same request and I assumed that he must have had the game scores already. I could not have been more wrong.

国际大师迈克·瓦尔沃(Mike Valvo)与特级大师罗恩·亨利(Ron Henley)进行了一年多的国际象棋训练。罗恩是阿纳托利·卡尔波夫的副手(国际象棋顾问)之一,作为一名投资交易员,他也是阿纳托利的财务顾问。当然,Mike Valvo 是 ACM 计算机国际象棋锦标赛的赛事总监,因此对 Deep Thought 非常熟悉。

International Master Mike Valvo had been doing some chess work with Grandmaster Ron Henley for over a year. Ron was one of Anatoly Karpov’s seconds (chess advisors) and, being an investment trader, he was also Anatoly’s financial advisor. Mike Valvo, of course, was the tournament director for the ACM Computer Chess Championships, and thus was quite familiar with Deep Thought.

与阿纳托利的比赛原定于 1990 年 2 月 2 日星期五在哈佛大学举行。比赛前一天晚上 9 点左右,迈克拜访了罗恩,为阿纳托利即将对阵荷兰大师扬·蒂曼的比赛提供了一台笔记本电脑,其中包含 ChessBase 数据库。当罗恩要求阿纳托利看深度思考游戏时,迈克感到很惊讶。事实证明,阿纳托利根本没有为“深思”比赛做好准备,尽管这场比赛很可能是他短暂的美国之行中最重要的一场比赛。

The match against Anatoly was supposed to take place on Friday 2 February 1990 at Harvard University. At about 9 pm on the night before the match, Mike visited Ron to deliver a laptop computer containing ChessBase databases for Anatoly’s upcoming match against the Dutch Grandmaster Jan Timman. Mike was surprised when Ron requested Deep Thought games for Anatoly to look at. It turned out that Anatoly had not prepared for the Deep Thought match at all, even though the match was very likely the most important event during his short trip to the United States.

我第一次见到加里时,他声称只有他和阿纳托利才是真正的“职业”棋手。嗯,与加里相比,阿纳托利是一个完全不同类型的职业球员。加里研究了 Deep Thought 的所有公开游戏及其弱点,包括开局。阿纳托利在与“深思”比赛之前,只是进行了一些肤浅的研究。

The first time I met Garry, he made the statement that only he and Anatoly were the true “professional” chess players. Well, compared to Garry, Anatoly was a quite different type of professional player. Garry studied all of Deep Thought’s public games and its weaknesses, including its opening. Anatoly did not do anything more than a superficial study before his match with Deep Thought.

当比赛安排好时,深思团队很惊讶这将是一场单场比赛。国际象棋比赛通常进行偶数局,以便双方玩家可以扮演相同数量的白人和黑人。单场比赛的传统是通过抛硬币或其他机会事件来选择比赛颜色。哈佛国际象棋俱乐部的组织者和我们自己都期望比赛前会随机选择颜色,特别是因为这将是一场单局比赛。事后看来,卡尔波夫阵营似乎很早就有不同的想法。

When the match was arranged, the Deep Thought team was surprised that it was going to be a one-game match. Chess matches are usually played with an even number of games so that both players get to play the same number of Whites and Blacks. The tradition for one-game matches was to select the playing color by a coin flip or some other chance event. The Harvard Chess Club organizers and ourselves all expected that the color would be chosen at random before the game, especially since it would be a one-game match. In hindsight, it seems that the Anatoly Karpov camp probably had a different idea early on.

以下是 IM Mike Valvo 撰写的有关这场比赛的文章的摘录:

Below is an excerpt from an article written by IM Mike Valvo on the match:

……然而,与费舍尔和卡斯帕罗夫不同的是,卡尔波夫满足于与黑方打平,而只用白方施压。

在我看来,卡尔波夫在看到机器对阿廖欣防御12的处理后,决定玩 1. e4 。我认为告诉他这台机器也可以玩 Caro-Kann 13是公平的,但我无意多说,因为我私下支持 Deep Thought。那时我大吃一惊,因为亨利现在打电话给丹尼·埃德尔曼并继续说服他应该给卡尔波夫白色棋子而不用抛硬币。爱德曼显然反对这个建议,但亨利却很有说服力。

… Unlike both Fischer and Kasparov, however, Karpov is content to play for a draw with Black and only press with White.

It seemed to me that Karpov had made up his mind to play 1. e4 after seeing the machine’s handling of the Alekhine Defense12. I thought it was only fair to tell him that the machine could also play the Caro-Kann 13, but I had no intention of saying anything more as I was privately rooting for Deep Thought. I was taken aback at that point because Henley now phoned Danny Edelman and proceeded to convince him that Karpov should be given the white pieces without a coin flip. Edelman was clearly resistant to the suggestion, but Henley was quite persuasive.

听到这个消息时,默里和我都惊呆了。作为多年的世界冠军,阿纳托利已经习惯了特殊待遇。如果他在比赛邀请时要求扮演怀特,我不会感到任何疑虑。当时我们可能不同意参加比赛,但也完全有可能同意参加比赛。阿纳托利最后一刻的要求留下了非常糟糕的印象。如果我们早点知道我们会对阵黑方,我们就可以为 Deep Thought 准备一个更好的开局。在高水平国际象棋中,良好的开局准备至关重要。提前充分了解您的颜色与特定的对手交手可能是胜负的区别。我们对事态的发展非常不满,但最终还是同意了这个要求,原因有两个。首先,我们不想大吵大闹。哈佛学生团体为这次活动付出了很多努力,我们感到有义务帮助他们取得成功。其次,我们不再仅仅代表我们自己。如果完全由我们决定,我们可能会说:“再见,很高兴见到你,卡尔波夫先生。” 一切都发生得太快,我们必须做出快速判断,而被迫下黑似乎是我们最不担心的。

Murray and I were stunned when we heard the news. Having been a World Champion for many years, Anatoly was used to special treatment. Had he asked to play White at the time of the match invitation, I would have felt no qualms about it. It was possible that we might not have agreed to play the match then, but it was also perfectly possible that we might. Anatoly’s last minute demand left a very bad taste. Had we known that we would be playing the Black side earlier, we could have prepared a much better opening for Deep Thought. At high level chess, good opening preparation is crucial. Knowing well in advance which color you will play against a specific opponent can be the difference between winning and losing. We were very unhappy about the turn of events, but we agreed to the demand in the end, for two reasons. First, we did not want to make a scene. The Harvard student group had put in a lot of work to make the event happen and we felt obliged to help them make it a success. Second, we were no longer only representing ourselves. Had it been entirely up to us, we might have said, “Good bye, nice meeting you, Mr. Karpov.” Everything happened too fast, we had to make a snap judgment, and being forced to play Black seemed to be the least of our worries.

阿纳托利看起来与我的预期大不相同。大多数描述都把他描绘成一个看起来相对精致的人。1990 年 2 月 2 日,当我第一次见到他时,他看起来很健康,尽管旅途中有点累。阿纳托利在比赛前不久就再婚了,婚姻生活似乎很适合他。

Anatoly looked quite different from what I expected. Most descriptions portrayed him as someone who looked relatively delicate. When I saw him for the first time on 2 February 1990, he appeared quite healthy, although a little bit tired from his trip. Anatoly had remarried shortly before the match, and married life seemed to agree with him.

在阿纳托利致开幕词后,比赛正式开始。正如迈克·瓦尔沃(Mike Valvo)推测的那样,阿纳托利(Anatoly)下出了 1. e4,即国王兵开局。在加里·卡斯帕罗夫(Garry Kasparov)和阿纳托利·卡尔波夫(Anatoly Karpov)的比赛之间,Deep Thought 经历了大卫·利维(David Levy)的比赛和一场电脑国际象棋锦标赛,期间肯·汤普森(Ken Thompson)为我们提供了他的卡罗·卡恩(Caro Kann)的开篇书。迈克·瓦尔沃(Mike Valvo)意识到了这一安排,因此他在之前的会面中考虑是否要告诉阿纳托利有关卡罗·坎恩(Caro Kann)的事情。碰巧的是,在两次介入的事件中,我们的对手都没有玩 1. e4,因此 Deep Thought 没有在公开游戏中玩 1.… c6,即 Caro Kann,对阵 1. e4。当我在棋盘上进行卡罗·卡恩防守时,阿纳托利没有表现出任何明显的情绪。与加里不同,阿纳托利的面部表情并不那么容易解读。

After an opening speech by Anatoly, the game was under way. As Mike Valvo surmised, Anatoly played 1. e4, the king pawn opening. Between the Garry Kasparov and the Anatoly Karpov matches, Deep Thought had gone through the David Levy match and one computer chess tournament, during which Ken Thompson had provided us with his opening book for the Caro Kann. Mike Valvo was aware of this arrangement, and hence his deliberation about whether to tell Anatoly about the Caro Kann during their earlier meeting. As it happened, in the two intervening events none of our opponents played 1. e4, so there was no public game where Deep Thought played 1.… c6, the Caro Kann, against 1. e4. Anatoly did not show any visible emotion when I made the Caro Kann defense on the chessboard. Unlike Garry, Anatoly’s facial expression was not as easy to read.

这场比赛的时间控制比对阵加里的比赛更快。阿纳托利·卡尔波夫的比赛​​每位选手进行一小时,而加里·卡斯帕罗夫的比赛每位选手进行 90 分钟。令人惊讶的是,在第16步“深思”有机会做出牺牲,这会让阿纳托利承受巨大的压力。如果比赛是按照常规的时间控制来进行的,那么 Deep Thought 就会做出牺牲。假设我们已经让它能够为比赛工作,那么与加里一起比赛的程序的 3 工作站版本可能会在比赛时间内找到牺牲。这是令人惊奇的是,一个微小的差异竟然可以极大地改变历史。

The game was played at a faster time control than in the match against Garry. The Anatoly Karpov game was played with one hour to each player, while the Garry Kasparov match was played with 90 minutes to each player. Surprisingly, by move 16 Deep Thought had a chance to make a sacrifice that would have placed Anatoly under tremendous pressure. Had the game been played with a regular time control, then Deep Thought would have made the sacrifice. The 3-workstation version of the program that played Garry potentially could find the sacrifice within the game time, assuming that we had got it working for the match. It is amazing how a minor difference could have changed the history dramatically.

由于“深思”无法在棋盘上找到牺牲,因此它下了第二好棋,阿纳托利获得了轻微的优势。第27手,深思下了一招乍一看相当难看的棋,引得观众哈哈大笑。赛后,阿纳托利评价说,那个丑陋的举动是唯一的好举动。我无法说出这样或那样的情况。我在比赛中注意到的只是,所下的丑陋的棋子比看起来更自然的棋子14好百分之一。

Since Deep Thought could not find the sacrifice over the board, it played the second best move and Anatoly obtained a slight edge. On the 27th move, Deep Thought played a move that at first glance looked quite ugly and the audience laughed. After the game, Anatoly commented that that ugly move was the only good move. I could not tell one way or another. All I noticed during the game was that the ugly move played was one hundredth of a pawn better than a more natural looking move14.

不幸的是,幸运女神并没有在深思身边停留太久。游戏进入车结局,双方都有一个王、一个车和一堆棋子。车的结局是客观制定的。比赛在很多回合中都保持平局,而深思一度有机会通过重复获得平局,但它拒绝了平局,因为它错误地评估了局面对自己有利。阿纳托利作为残局高手,知道得更清楚,他在深思不理解的位置找到了一些资源。Deep Thought 在 65 步后最终输掉了比赛。比赛结束时,阿纳托利已经到了最后一刻,而深思还有大约二十分钟的时间。考虑到卡尔波夫剩​​下的时间有限,当我辞去《深思》的职务时,观众们有点惊讶。在我看来,这是一场表演赛,阿纳托利显然知道自己在做什么,浪费大家的时间强迫他把比赛打到最后是没有意义的。

Unfortunately, Lady Luck did not stay with Deep Thought very long. The game entered into a rook ending, where each side had a king, a rook and a bunch of pawns. The rook ending was objectively drawn. The game remained drawn for many moves, and at one point Deep Thought had a chance for a draw by repetition, but it spurned the draw as it erroneously assessed the position to be in its favor. Anatoly, being the virtuoso endgame player, knew better and he found some resources in the position that Deep Thought did not understand. Deep Thought ended up losing the game after 65 moves. Anatoly was down to his last minute when the game ended, while Deep Thought still had about twenty minutes. The audience was a little bit surprised when I resigned for Deep Thought, given the limited time that Karpov had left. In my view, it was an exhibition game, Anatoly obviously knew what he was doing, and there was no sense wasting everybody’s time to force him to play the game to the bitter end.

赛后,迈克·瓦尔沃对观众表示,这场比赛可以被视为Deep Thought的道德胜利。我认为这并不是一场空洞的道德胜利,而是世界顶级棋手之一给我上的一堂精彩的国际象棋课。我们需要对 Deep Thought 或其后继者的残局玩法做一些扎实的工作。《Deep Thought II》的设计很大程度上受到了卡尔波夫游戏的影响,很少有游戏对《Deep Thought》或其后继者的设计产生类似的影响。总而言之,阿纳托利·卡尔波夫的国际象棋能力必须得到尊重。如果让卡尔波夫下白棋是教训的代价,那么这笔交易还算不错。

After the game, Mike Valvo said to the audience that the game could be viewed as a moral victory for Deep Thought. I viewed it not so much as a hollow moral victory, but rather a good chess lesson given by one of the top chess players in the world. We needed to do some solid work about Deep Thought’s or its successor’s endgame play. The design of Deep Thought II was influenced quite substantially by the Karpov game, and there were very few games that had similar effects on the design of Deep Thought or its successors. When all was said and done, Anatoly Karpov’s chess ability had to be respected. If letting Karpov play White was the price for the lesson, it was not so bad a deal.

    嘿,夸张是免费的。

    Hey, hyperbole is free.

    好吧,如果你期待我的论文答辩有什么火花的话,我不得不说,我的论文答辩进行得很顺利。辩护向公众开放,柏林纳博士作为公众成员参加了辩护。虽然柏林博士可能不同意我所采取的指示,但他并没有给我带来任何麻烦。

    Well, if you were expecting any fireworks at my thesis defense, I have to say it went smoothly. The defense was open to the public and Dr Berliner attended it as a member of the public. While Dr Berliner might have disagreed with the directions that I was taking, he did not give me any trouble.

    国际象棋世界冠军挑战者的选择制度不时发生变化。在 20 世纪 90 年代中期之前的大部分历史中,挑战者都是通过一系列淘汰赛(也称为候选人赛)选出的。

    The system for selecting the challenger for the World Chess Champion changed from time to time. For most of the recent history up to the mid 1990s, the challenger was selected by a series of elimination matches also known as the Candidate matches.

    在深度思考量表上,一个棋子值 128 分。在深蓝等级中,一个棋子值 100 点。

    A pawn was worth 128 points on the Deep Thought scale. On the Deep Blue scale, a pawn was worth 100 points.

    在ChipTest、Deep Thought、Deep Thought II和Deep Blue中,棋盘上距离游戏位置最远的位置是由国际象棋硬件搜索的。这些机器使用在通用计算机控制下运行的专用国际象棋硬件。对于12层搜索,前八层可以由通用计算机上的软件搜索,后四层可以由国际象棋硬件直接搜索。

    In ChipTest, Deep Thought, Deep Thought II and Deep Blue, the positions most distant from the game position on the chessboard are searched by the chess hardware. These machines use special purpose chess hardware running under the control of a general purpose computer. For a 12-ply search, the first eight plies might be searched by the software on the general purpose computer, and the last four plies might be searched by the chess hardware directly.

    在操作电脑与人类玩家对战时,我会尽量保持一本正经的表情,但在与电脑对战时,没有必要隐藏我的真实感受。

    I tried to keep a poker face when operating the computer against human players but in games against computers, there was no need to hide my true feelings.

    Jonathan 将注意力转移到计算机跳棋上,最终与几位同事创建了一个名为 Chinook 的世界冠军级别跳棋程序。乔纳森的优秀著作《一跃领先:挑战跳棋中的人类霸权》描述了奇努克的故事

    Jonathan moved his attention to computer checkers, and eventually, with a few colleagues, created a World Championship level checkers program named Chinook. Chinook’s story is described in Jonathan’s excellent book One Jump Ahead: Challenging Human Supremacy in Checkers.

    即使加里没有打破开局,它很可能仍然会输掉比赛。

    Most likely it would still lose the game even if Garry had not busted its opening.

    1997 年深蓝比赛结束后,加里因财务问题解雇了安德鲁·佩奇。

    Garry fired Andrew Page after the 1997 Deep Blue match over financial issues.

   好吧,一天工作 10,000 美元是大多数人会不加考虑地承担的风险。

   Okay, $10,000 for one day’s work is a risk that most people would undertake without any consideration.

   Peter 对 Deep Thought 项目的主要贡献是选定棋子结局的残局数据库。他没有积极参与 Deep Thought 的编程。Peter 更喜欢使用国际象棋机器,在他的博士论文研究中广泛使用 ChipTest。当 Murray、Thomas 和我去 IBM 时,它就被留在了卡内基梅隆大学。

   Peter’s main contribution to the Deep Thought project was endgame databases for selected pawn endings. He did not get actively involved in the programming of Deep Thought. Peter was more a user of the chess machines, using ChipTest extensively in his PhD thesis research. It had been left behind at Carnegie Mellon when Murray, Thomas and I went to IBM.

   对于不懂国际象棋的人来说,阿廖欣防御通常被认为是黑棋的一种可疑开局。Deep Thought 在对抗防守较弱的选手时取得了不错的成绩,但阿纳托利·卡尔波夫 (Anatoly Karpov) 因击败敢于对他进行阿廖欣防守 (Alekhine Defense) 的顶级大师而闻名。

   For the chess uninitiated, the Alekhine Defense is generally considered a somewhat suspect opening for Black. Deep Thought had good results against weak players with the defense, but Anatoly Karpov had been known to mow down top grandmasters daring to play the Alekhine Defense against him.

   卡罗·卡恩 (Caro Kann) 是黑方相对稳健的防守。讽刺的是,这也是阿纳托利当时下黑棋时最喜欢的防守。

   Caro Kann is a relatively sound defense for Black. Ironically, it was also Anatoly’s favorite defense when he played Black at the time.

   国际象棋程序员通常不会在公开场合承认这一点,但这种情况却时常发生。“出色”的棋步通常只比失败的棋步“好”一点点。

   Chess programmers usually don’t admit it in public, but this happens all the time. The “brilliant” moves are frequently only “better” than the losing moves by the slightest difference.

 

 

国际商业机器公司

IBM


 

 

第 8 章

CHAPTER 8

我们需要一个新名字

We Need a New Name


文化冲击

A Cultural Shock

从卡内基梅隆大学熟悉的环境搬到位于纽约州威彻斯特县的 IBM TJ Watson 研究中心,我花了一段时间才适应。

The move from the familiar environment at Carnegie Mellon University to the IBM T. J. Watson Research Center in Westchester County, New York took me a while to get used to.

在我们加入 IBM Research 的几个月前,我们接到了兰迪·穆利克 (Randy Moulic) 的电话,他将成为我们的经理。Randy 对这个项目感到很兴奋,并希望在我们加入 IBM 之前到匹兹堡拜访我们。这次会议主要是社交性的。我原以为是一个相对正式、比我们年长得多的人,但兰迪结果很热情,只比我们大十岁左右。事实上,我的第一个心理反应是“我们的新经理有一张娃娃脸”。兰迪心地很年轻,有着乐观的人生观。选他当我们经理的人干得很好;他可能比大多数 IBM 经理更接近我们的思维方式。会议进行得很顺利,我以为在IBM的生活过渡会比较顺利。我低估了我即将受到的文化冲击。

A couple of months before we joined IBM Research, we got a call from Randy Moulic, who was to be our manager. Randy was excited about the project and wanted to visit us in Pittsburgh before we joined IBM. The meeting was mostly social. I was expecting someone relatively formal and much older than we were, but Randy turned out to be quite cordial, and only about ten years older. In fact, my first mental response was “our new manager has a baby face”. Randy was quite young at heart and had a sunny outlook on life. Whoever selected him to be our manager did a good job; he was probably closer to our way of thinking than most IBM managers. The meeting went well, and I thought that the transition to life in IBM would be relatively smooth. I underestimated the cultural shock that I was about to receive.

卡内基梅隆大学计算机科学研究生的生活很舒适。该部门的津贴充足,而且匹兹堡的生活成本相对较低。部门里的每个人都在某种程度上认识你,你也至少知道部门里每个人的名字。如果这是他们想要的,任何人都可以拥有很大的自主权和自由。如果我需要任何帮助,我知道该与谁交谈,或者至少知道可以向我指出合适的人。该部门每天 24 小时有效运作。即使在半夜,你总能找到一些同学,偶尔也能找到一些教员。

Life as a computer science graduate student at Carnegie Mellon was comfortable. The stipend from the Department was adequate, and the cost of living in Pittsburgh was relatively low. Everyone in the Department knew you to some extent, and you knew everyone in the Department at least by name. Anyone could have a great deal of autonomy and freedom if that was what they wanted. If I needed any help, I knew with whom to talk, or at least someone who might be able to point me to the right person. The Department effectively operated 24 hours a day. Even in the middle of the night, you could always find some fellow students or occasionally some faculty members about.

与匹兹堡相比,威彻斯特县的生活成本要高得多——超市的食物大约高出 10% 到 20%,而我的新公寓的租金大约是匹兹堡的三倍。幸好我的新工资比我以前的学生津贴高得多。标价带来的冲击需要一些时间来适应,但这并不像工作环境的改变那么难。

Compared to Pittsburgh, Westchester County had much higher living cost—the food at the supermarket was about 10–20 percent higher, and the rent for my new apartment was about three times more than in Pittsburgh. It was a good thing that my new salary was much higher than my old student stipend. The sticker shock took some time to get used to, but it was not as hard as the change in the work environment.

卡内基梅隆大学的移民课程在帮助新研究生融入该系的新环境方面做得很好。IBM 研究中心没有类似的流程来帮助新人。我在 IBM 的头几个月都在处理看似永无休止的文书工作,并试图适应 IBM 的计算环境,这与典型大学的计算环境有很大不同1。此外,我不再知道周围的人在做什么。每个人似乎都忙着自己的项目,与自己小组之外的交流很少。大多数人似乎也按正常时间工作。通常到晚上 7 点,大楼就几乎空了。

The Immigration Course at Carnegie Mellon did a good job of easing new graduate students into the new environment at the Department. IBM Research did not have a comparable process to help newcomers. The first few months of my life at IBM were spent dealing with the seemingly never-ending paperwork, and trying to cope with the IBM computing environment, which was quite different from the computing environment in typical universities1. Furthermore, I no longer knew what the other people around me were doing. Everyone seemed to be busy with his own project, and there was very little communication outside one’s own group. Most of the people also appeared to work regular hours. The building was usually almost empty by 7 o’clock in the evening.

然而,最有趣的新体验是关于程序名称的所有讨论。当我选择“Deep Thought”这个名字时,它似乎适合击败国际象棋世界冠军的机器。我不知道有些人可能会遇到问题。潜在麻烦的第一个迹象出现在 Deep Thought 参加的第一场计算机国际象棋锦标赛上,甚至在我们开始与 IBM 进行就业谈判之前。那是 1988 年 ACM 电脑国际象棋锦标赛。阿尔伯塔大学教授托尼·马斯兰 (Tony Marsland) 在锦标赛期间组织了一场小组会议,我是小组成员之一。由于某些原因,托尼在介绍我时至少用了三四次“深喉”这个名字,但每次都立即纠正自己。“深喉”是《华盛顿邮报》水门事件调查记者线人的代号,也是一部你不会带你妈妈去看的电影的名字。

However, the most amusing new experience was all the discussion about the name of the program. When I chose the name Deep Thought, it seemed to be appropriate for the machine that would defeat the World Chess Champion. I had no idea that some people might have problems with it. The first hint of potential trouble came at the very first computer chess tournament that Deep Thought played in, before we even started the employment talk with IBM. It was at the 1988 ACM Computer Chess Championship. Tony Marsland, a professor from the University of Alberta, organized a panel session alongside the Championship and I was one of the panelists. For some reasons, when introducing me, Tony used the name “Deep Throat” at least three or four times although immediately correcting himself each time. “Deep Throat” was the code name of the informer for the Watergate investigative reporters from the Washington Post, and it was also the name of a movie that you would not take your mother to see.

加里·卡斯帕罗夫 (Garry Kasparov) 与 Deep Thought 的比赛结束后,IBM 内部和外部对 Deep Thought 这个名字的抱怨纷至沓来。看来有很多人可以当他们听到“Deep Thought”这个名字时,不禁想起“Deep Throat”。情况变得更糟了。默里·坎贝尔的妻子吉娜在一所天主教大学工作。学院院长是一位年迈的修女,在与吉娜交谈时,修女一直使用“深喉”这个名字。休斯顿,我们有一个问题。

After the Garry Kasparov vs. Deep Thought match, complaints of the Deep Thought name poured in from both inside and outside IBM. It appeared that there were a lot of people who could not help thinking about “Deep Throat” when they heard the name Deep Thought. It got much worse. Murray Campbell’s wife Gina was working at a Catholic college. The President of the college was an elderly nun, and during a conversation with Gina, the Sister kept on using the name “Deep Throat”. Houston, we have a problem.

在卡内基梅隆大学,我们选择了自己喜欢的名字。我们在 IBM 无法做到这一点。鉴于 IBM 是一家大公司,我们必须小心不要侵犯他人的商标。让更多 IBM 研究人员参与命名过程似乎也是个好主意,因此通信部宣布为未来的 IBM 国际象棋机器命名竞赛。在提交的参赛作品中,“Blue Chips”这个名字是我最喜欢的,但一家计算机公司已经在使用这个名字了。其他几个名字也有同样的问题。最终,我们选择了“深蓝” 2这个名字,由彼得·布朗(Peter Brown)提出,他是我们被IBM聘用的部分责任的同班同学。为了赢得比赛,彼得与我们一起享用了一顿丰盛的晚餐,并有机会玩 Deep Thought。他拒绝了奖金的第二部分3

At Carnegie Mellon, we chose whatever name we liked. We could not do that at IBM. Given that IBM was a big corporation, we had to be careful not to infringe on someone else’s trademark. It also seemed a good idea to get more IBM researchers involved in the naming process, so the Communications Department announced a contest for the name of the future IBM chess machine. Of the submitted entries, the name “Blue Chips” was my favorite, but a computer company was already using that name. A few other names also had the same problem. In the end, the name “Deep Blue”2 was chosen, submitted by Peter Brown, that same classmate who was partially responsible for our being hired by IBM. For winning the contest, Peter was treated to a nice dinner with us, and a chance to play Deep Thought. He declined the second part of his award 3.

短期观点

The Short-term View

当我们 1989 年加入 IBM 时,它是世界上最赚钱的公司之一,直到 1993 年左右的财政年度都非常好。作为一名研究生,我习惯于以最少的支出进行运营。我们对 ChipTest 的预算为零,对 Deep Thought 4的预算为 5000 美元。当你的预算像我们一样紧张时,你最好节俭一点。对我来说,考虑到项目早期阶段大部分关键工作都在硬件方面,IBM 同意立即雇用一个完整的团队已经是一个很大的惊喜。所以当我们的经理 Randy Moulic建议我们聘请一位国际象棋大师作为顾问,我的第一反应是“为什么?”。我的感觉是,我们确实需要大师的帮助来进行最后的推动,但我们还没有真正准备好与他们合作,因为新硬件仍然只是一个闪光。

When we joined IBM In 1989, it was one of the most profitable companies in the world, with very good financial years until 1993 or so. As a graduate student, I was used to operating with the minimum amount of outlay. We had a nil budget for ChipTest, and a $5000 budget for Deep Thought4. You had better be frugal when you have a budget as tight as we did. To me, it was already a big surprise that IBM agreed to hire a full team immediately, considering that most of the critical work would be in hardware during the early phases of the project. So when our manager Randy Moulic suggested that we hire a Chess Grandmaster as a consultant, my immediate gut reaction was “What for?”. My feeling was that we would indeed need help from Grandmasters for the final push, but we were not really ready to work with them yet, since the new hardware was still no more than a glint in the eye.

我没有意识到我们不能再只着眼于长远。回到卡内基梅隆大学,“深度思考”并不是一个官方项目,教员们对我们的表现并没有抱有很高的期望。在加入 IBM 之前,我们有自己设定的里程碑要实现,但我们不必在计算机国际象棋世界中保持持续存在。理论上,我们可以连续多年回到绘图板,而不出现在任何计算机国际象棋比赛中。我们在 IBM 却没有这种奢侈。我们的到来是一场高调的活动。两个月内,我们因与加里·卡斯帕罗夫的表演赛而登上了《纽约时报》和《华尔街日报》的头版。如此高的知名度是有代价的。我们突然必须定期在公开场合竞争并获胜。提供短期绩效成为我们的一个重要关注点,即使它可能与项目的长期进展相冲突。很可能正是这种担忧促使兰迪建议聘请一位大师。

What I did not realize was that we could no longer afford to take only a long-term view. Back at Carnegie Mellon, Deep Thought was not an official project and the faculty did not have high expectations about how well we would do. Before we joined IBM, we had our own self-imposed milestones to meet, but we did not have to maintain a constant presence in the computer chess world. In theory, we could go back to the drawing board for years on end without showing up at any computer chess event. We did not have this luxury at IBM. Our arrival there was a high-profile one. Within two months, we were on the front page of the New York Times and the Wall Street Journal as a result of the exhibition match with Garry Kasparov. Such a high profile came with a price. We suddenly had to compete regularly in public and win. Providing short-term performance became an important concern for us even if it might be in conflict with the long-term progress of the project. It was most likely this concern that drove Randy to suggest hiring a Grandmaster.

大师马克西姆·德鲁吉(Maxim Dlugy)出生于俄罗斯,但他在很小的时候就来到了美国,并在美国接受了大部分国际象棋教育。在某种程度上,马克西姆是美国和俄罗斯文化的交叉产物。当我们与一些国际象棋朋友交谈时,提到了他的名字,我们邀请他来 IBM 研究院参观。他善于表达,风度翩翩,经过一些内部讨论,他被聘为国际象棋顾问。穆雷作为球队的国际象棋专家,将是与他合作的主要人员。

Grandmaster Maxim Dlugy was born in Russia, but he came to the United States when he was quite young and received most of his chess education in the US. To some extent, Maxim was a cross product of both the American and the Russian cultures. His name came up when we talked with some of our chess friends and we invited him to come up to IBM Research for a visit. He was articulate and personable, and after some internal discussions he was hired as a chess consultant. Murray, being the chess expert of the team, would be the main person working with him.

当马克西姆加入时,我们对他没有明确的责任。在最初的几周里,他会进来,与 Deep Thought 进行几场快速比赛,大部分都输了,然后发表一些评论。随后,他会引入一些测试位置,其中一些来自他最近的锦标赛比赛,看看 Deep Thought 会提出什么建议。有时,令深思懊恼的是,他会发现自己错过了彻底赢得比赛的机会。Maxim 的评论可能非常有用,但如果没有至少部分设计新硬件,我们就无法利用它们。

When Maxim came on board, we had no clear responsibilities for him. In the first few weeks, he would come in, play a few quick games against Deep Thought, losing most of them, and then make a few comments. Later on, he would bring in some test positions, some from his recent tournament games, and see what Deep Thought came up with. Occasionally, much to his chagrin, Deep Thought would find that he had missed chances to win the games outright. The comments from Maxim were potentially very useful, but we were not in a position to make use of them without the new hardware at least partially designed.

Deep Thought的开局书虽然达到了宗师级别,但并没有达到宗师标准。在从马克西姆那里得到有用但不能立即可用的工作后,我们请他写一本更好的入门书。在这种新的工作方式中,马克西姆首先会带来他在家准备的一本开篇书片段。Deep Thought会检查开卷中的所有错误位置,Murray会筛选明显的错误,然后Maxim会检查标记的错误并提出替代方案。

Deep Thought had an opening book that was not up to Grandmaster standard, even though it was playing at Grandmaster level. After getting useful but not immediately usable work from Maxim, we asked him to work on a better opening book. In this new way of working, Maxim would first bring in an opening book fragment that he prepared at home. Deep Thought would check out all the positions for errors in the opening book, Murray would screen for the obvious errors, and then Maxim would check the flagged errors and come up with alternatives.

然而,《德鲁吉》的开篇却从未被完全检查过。1991 年,当美国信孚银行招募顶级棋手作为潜在的股票或货币交易员时,Maxim 离开了 IBM,加入了他们5我们在 Deep Thought 对阵德国国家队的第 6 场比赛中使用了他的开场白,结果好坏参半。这本开局书的部分内容直接来自国际象棋开局百科全书(ECO)。ECO主要是大师作品的集合,据说其中一些大师在他们所写的部分中故意设置了陷阱。对于国际象棋棋手来说,众所周知,ECO 不能完全信任。在与德国选手的一场比赛中,Deep Thought 直接陷入了开局陷阱。ECO评估这个位置对Deep Thought来说是可以玩的,但实际上却是惨败。如果对该仓位进行检查,我们会注意到 Deep Thought 将其评估为完全典当,没有任何仓位补偿,这通常是该仓位丢失的良好迹象。

The Dlugy opening book, however, was never fully checked. In 1991, when Bankers Trust was recruiting top chess players as potential equity or currency traders, Maxim left IBM and joined them5. We used his opening book in a Deep Thought vs. the German national team match6, with mixed results. Part of this opening book came straight out of the Encyclopedia of Chess Openings (ECO). The ECO is largely a collection of works from Grandmasters and some of the them were said to have deliberately laid opening traps in the sections that they wrote. Among chess players, it is well known that the ECO cannot be fully trusted. In one of the games against the German players, Deep Thought walked right into an opening trap. ECO assessed the position as playable for Deep Thought, but actually it was a dead loss. Had the position been examined, we would have noticed that Deep Thought assessed it as a full pawn down without any positional compensation, which is usually a good indication that the position is lost.

Dlugy 的实验基本上不成功。当 Deep Thought II 全面投入使用时,他的开篇书籍被降级为备用角色。到 Deep Blue 时代,大部分 Dlugy 分析已不再活跃。然而,我们确实从这个过程中学到了一些东西——如何不与大师合作。为了与他们正确合作,我们需要设定明确的目标并建立一些衡量我们进展的方法。

The Dlugy experiment was largely unsuccessful. His opening book was relegated to a backup role when Deep Thought II became fully operational. By the time of Deep Blue, most of the Dlugy analysis was no longer active. However, we did learn something from the process—how not to work with a Grandmaster. To work with them properly, we needed to set clear goals and establish some means to measure our progress.

深思二和更美好的牧场

Deep Thought II and the Greener Pasture

Deep Thought II,也称为 Deep Blue Prototype,在 1991 年至 1995 年的职业生涯中发挥了多种有用的作用。它是最强大的国际象棋计算机,直到 1996 年 Deep Blue 取代它。回想起来,构建 Deep Thought II 的决定可能是一个失误。由于Deep Thought II,深蓝推迟了大约两年,尽管这一推迟的部分原因是不可预见的人事变动。从好的方面来说,如果没有 Deep Thought II,对深蓝成功至关重要的 Joe Hoane 可能不会加入该团队。

Deep Thought II, also known as Deep Blue Prototype, served several useful purposes during its career from 1991 to 1995. It was the strongest chess computer until Deep Blue supplanted it in 1996. The decision to build Deep Thought II, in retrospect, was probably a mistake. Because of Deep Thought II, Deep Blue was delayed by about two years, although part of this delay was due to unforeseen personnel changes. On the plus side, without Deep Thought II, Joe Hoane, who became critical to the success of Deep Blue, might not have joined the team.

构建 Deep Thought II 的最初决定似乎是一个不错的决定。Deep Thought 仍然比最接近的计算机国际象棋竞争对手强约 200 个评级点,并且在与最好的对手的每四场比赛中获胜三场。但主要计算机国际象棋赛事的瑞士锦标赛形式意味着它赢得比赛的机会只有百分之五十左右,尽管它始终是最受欢迎的。当我们还是研究生时,赢得计算机国际象棋比赛的胜利是令人欣慰的,但除此之外并不重要。在企业环境中,尽管统计数据对我们不利,但我们仍然面临着继续获胜的巨大政治压力。官方称,构建 Deep Thought II 的原因是为了将其用作探索并行搜索算法的原型。但主要原因实际上是为了在我们的计算机竞争对手和我们之间拉开很大的差距,以便在我努力设计新的国际象棋芯片时,该项目可以持续很长一段时间。

The initial decision to build Deep Thought II seemed to be a good one. Deep Thought was still about 200 rating points stronger than the nearest computer chess competitors, and would win three out of every four games against the best opponents. But the Swiss tournament format of the major computer chess events meant that its chance of winning an event was only about fifty percent, even though it was always the heavy favorite. When we were just graduate students, winning computer chess events was gratifying but otherwise unimportant. In a corporate environment, there was tremendous political pressure to keep on winning, even though the statistics were stacked against us. Officially, the reason for building Deep Thought II was to use it as a prototype for exploring parallel search algorithms. But the main reason was really to open up a large gap between our computer competitors and us, so that the project could go on for an extended period while I struggled through the design of the new chess chip.

在我看来,构建 Deep Thought II 还有另一个原因。我想强迫托马斯·阿南塔拉曼重写国际象棋机的程序。

From my point of view, there was another reason to build Deep Thought II. I wanted to force Thomas Anantharaman to rewrite the program for the chess machine.

托马斯和我走了很远的路。他是除了我之外第一个研究 Deep Thought 系列国际象棋机器的人。我们已经非常接近了,但当我们加入 IBM 时,我们已经就如何对机器进行编程进行了长期的争论。在进入 ChipTest 之前,Thomas 编写了一个纯软件国际象棋程序。为了加速他的程序,他违反了许多良好编程的规则,特别是他的程序充满了许多“goto”语句,即分支到程序中另一个地方的计算机命令。这是一个经典的“意大利面条”代码——一个控制流程错综复杂的程序,普通人无法轻易遵循。

Thomas and I went back a long way. He was the first person besides me to work on the Deep Thought series of chess machines. We were quite close, but by the time we joined IBM we had been in a long running debate on how to program the machine. Prior to his ChipTest days, Thomas had written a software-only chess program. In order to speed up his program, he violated many of the precepts of good programming and, in particular, his program was riddled with many “goto” statements, computer commands to branch to another place in the program. It was a classic “spaghetti” code—a program with convoluted, intricate control flow that could not be easily followed by normal human beings.

托马斯不再需要编写国际象棋程序当 ChipTest 硬件开始工作时,采用“意大利面条”编码风格。但由于他只是将其作为一种爱好,我无法真正强迫他做出改变。乞丐不能挑剔。即使对于 Thomas 来说,ChipTest 软件也变得非常笨拙,当 Deep Thought 构建时,他最初同意以团队其他成员可以理解的方式重写软件。但部分是因为我延迟交付 Deep Thought 硬件,也可能是因为 Thomas 与他的论文研究相关的工作量,Deep Thought 软件是根据 ChipTest 软件进行修改的,并保留了其所有“意大利面条”代码。

It was no longer necessary for Thomas to write chess programs in the “spaghetti” coding style when the ChipTest hardware started working. But since he was doing it as a hobby, I could not really force him to make the change. Beggars cannot be choosers. The ChipTest software became sufficiently unwieldy even for Thomas, and when Deep Thought was being built, he initially agreed to rewrite the software in a manner that the rest of the team could understand. But partly because of my delay in delivering the Deep Thought hardware, and possibly partly because of Thomas’ work load related to his thesis research, the Deep Thought software was modified from the ChipTest software, and retained all of its “spaghetti” code.

Deep Thought 获得第二届 Fredkin 中级奖后,Thomas 决定放弃原来的论文题目,将 Deep Thought 上的工作变成论文,但“意大利面条”依然存在。当我们加入 IBM 时,Deep Thought 已经有大约 100,000 行代码,大部分都是“意大利面条”编码风格,甚至 Thomas 也很难跟踪软件的工作原理。Deep Thought II 比 Deep Thought 拥有更多的国际象棋处理器(最多 24 个),因此需要全新的软件,而 Thomas 终于有充分的理由来更改程序了。

After Deep Thought won the Second Fredkin Intermediate Prize, Thomas decided to drop his original thesis topic and turn his work on Deep Thought into a thesis, but the “spaghetti” remained. By the time we joined IBM, Deep Thought had about 100,000 lines of code, mostly in the “spaghetti” coding style, and even Thomas had a hard time keeping track of how the software worked. Deep Thought II, with many more chess processors (up to twenty-four) than Deep Thought, would require brand new software, and Thomas would finally have a good reason to change the program.

事情并没有像我希望的那样发展。

Things did not work out as I hoped.

当我们开始 Deep Thought II 的设计时,Randy 聘请了一名合同工程师与我一起进行硬件设计。他最初的工作非常出色,让我可以将大部分时间花在新的国际象棋芯片上,而不是监督新机器的设计。在完成工作之前,这位合同工程师离开 IBM 外部工作。突然间,我不得不接管整个设计。这次转移占用了芯片工作几个月的时间。

When we started the design for Deep Thought II, Randy hired a contract engineer to work on the hardware design with me. He did a creditable job initially, freeing me to spend most of my time on the new chess chip instead of overseeing the design of the new machine. Before completing his work, the contract engineer left for a job outside IBM. Suddenly, I had to take over the whole design. This diversion took several months away from the chip work.

还有更多的问题接踵而至。

There were more problems to come.

当我们都是卡内基梅隆大学的研究生时,托马斯在演讲小组工作。多年来,他在我进行芯片设计工作的同一台机器上运行两个名为“ex”和“lpc”的程序。我一直认为这些程序是他语音研究的一部分,特别是“lpc”程序听起来像“线性预测编码”程序,这是语音识别中使用的一种行之有效的方法,对于语音识别来说它非常有意义他去努力。当我即将离开卡内基梅隆大学时我才发现他的两个项目都与他的演讲工作无关。“ex”程序原来是一个股票程序交易所7,程序“lpc”用于预测股票市场的走势。托马斯长期以来一直对股票市场感兴趣。

When we were both graduate students at Carnegie Mellon, Thomas was working in the Speech group. For many years, he was running two programs, named “ex” and “lpc”, on the same machine on which I was doing my chip design work. I had always assumed that the programs were part of his speech research—in particular, the “lpc” program sounded like a “Linear Predictive Coding” program, which is a well-established method used in speech recognition, and it made perfect sense for him to work on it. I only found out when I was about to leave Carnegie Mellon that neither of his programs was related to his speech work. The program “ex” turned out to be a program for the stock exchange7, and the program “lpc” was used to predict the movement of the stock market. Thomas had had an interest in the stock market for a long while.

当华尔街公司与安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 接洽并提出在硅谷成立一家初创公司的计划时,他邀请托马斯为这家初创公司兼职工作。安德烈亚斯最终拒绝了这笔交易,但托马斯对此很感兴趣,因此他继续与该公司进行谈判。一段时间后,他达成了一项协议,认为与他在 IBM 的工作不存在利益冲突。不幸的是,IBM律师不同意,因为该交易涉及知识产权转让。托马斯最终离开 IBM 前往华尔街,新的薪水是他原来收入的两倍多。

When Andreas Nowatzyk was approached by the Wall Street firm with the plan of forming a startup in Silicon Valley, he asked Thomas to work for the startup on a part-time basis. Andreas eventually turned down the deal but Thomas was sufficiently interested that he continued his own negotiations with the firm. After some time, he reached a deal, believing that there was no conflict of interest with his IBM job. Unfortunately, IBM lawyers disagreed, because the deal involved the transfer of intellectual property. Thomas ended up leaving IBM for the greener pasture of Wall Street, with a new salary more than twice what he had earned.

在托马斯离开之前,我正在考虑创建我自己版本的国际象棋程序。我对提高程序的搜索选择性有一些新的想法,也就是说,让程序沿着“有趣”的路线进行比 Deep Thought 以前所做的更深入的搜索。我希望看到这些想法得以实现8但考虑到我的硬件设计工作,我并没有真正认真地自己编写国际象棋程序。我只是在说话,希望托马斯能够接受这些想法并付诸实践,就像我们回到卡内基梅隆大学时他经常做的那样。托马斯去世后,我被迫亲自编写一个新的国际象棋程序9

I was tinkering with the idea of creating my own version of the chess program before Thomas left. I had some new ideas for improving the search selectivity of the program, that is, making the program search much more deeply along “interesting” lines than Deep Thought ever had done. I wanted to see the ideas implemented8 but, given my hardware design work, I was not really serious about writing the chess program myself. I was just talking and hoping that Thomas would pick up the ideas and run with them as he had done frequently when we were back at Carnegie Mellon. With Thomas now gone, I was forced to write a new chess program for real, myself9.

亚瑟·约瑟夫·霍恩 (Arthur Joseph Hoane) 是 IBM 研究中心的一名高级程序员,我们称之为 Joe,他之前曾表达过他对国际象棋项目的兴趣。Thomas 离开后,我们的员工人数增加了10 人,我们需要有人来帮助编写 Deep Thought II 程序。乔毕业于伊利诺伊大学厄巴纳-香槟分校,我们知道他是一位上进心强的优秀程序员。“深蓝”不是一个普通的 IBM 项目;它是 IBM 的一个项目。这是一个移植的项目来自 IBM 外部,并且团队成员具有上进心。在团队加入IBM之前,我们明确告诉公司,除非有机会打造“终极”棋机,否则我们不会加入。与典型的 IBM 项目不同,深蓝项目的经理扮演的是支持角色,而不是项目领导者。事实上,当时并没有项目负责人。团队成员能够独立完成事情至关重要。经过一番讨论和大量文书工作后,乔被欢迎加入该小组。然而,他还没有准备好接替托马斯,所以我没有休息的机会。我仍然需要为 Deep Thought II 编写新的并行搜索代码和新的选择性搜索代码。

Arthur Joseph Hoane, a staff programmer at IBM Research whom we called Joe, had previously expressed his interest in the chess project. With Thomas gone, we had an extra head count10 and we needed someone to help program Deep Thought II. Joe was a graduate of the University of Illinois at Urbana-Champaign, and we knew him to be a self-motivated, excellent programmer. Deep Blue was not an ordinary IBM project; it was a project transplanted from outside IBM, and with team members who were self-motivated. Before the team joined IBM, we explicitly told the company that we would not join unless we were given the chance to build the “ultimate” chess machine. Unlike typical IBM projects, the manager of the Deep Blue project was filling a support role, and not the project leader. As a matter of fact, there was no project leader. It was critical that members of the team be able to do things on their own. After some discussion and lots of paperwork, Joe was welcomed into the group. However, he was not yet ready to step into Thomas’ shoes, so there was no break for me. I still had to program the new parallel search code and the new selective search code for Deep Thought II.

在卡内基梅隆大学,我为论文中描述的并行搜索算法编写了一个模拟器。使用这个模拟器作为新程序的起始模板,我度过了更轻松的时光。然而,这是一个缓慢的过程。编写软件从来都不是我要做的有趣事情的前列。

At Carnegie Mellon, I wrote a simulator for the parallel search algorithms described in my thesis. With this simulator as the starting template for the new program, I had an easier time. However, it was a slow process. Writing software was never high on my list of interesting things to do.

Deep Thought II 拥有多达 24 个国际象棋处理器,新的国际象棋程序旨在使用所有这些处理器。该程序的第一个版本没有实现 Deep Thought 中使用的任何选择性扩展,也没有实现我的搜索想法。花了几个月的时间才完成。新程序比 Deep Thought 快大约 5 到 8 倍。Thomas在论文中声称,Deep Thought中使用的搜索扩展仅在性能上产生了很小的改进,这不足以弥补Deep Thought II所享有的5-8倍的速度优势。然而,Deep Thought 与 Deep Thought II 的第一个版本在正面竞争中处于平局。Maxim Dlugy 当时已经前往 Bankers Trust,但他还有一些 IBM 合同时间需要填补,并且他与 Deep Thought II 进行了一场比赛。赛后他认为《Deep Thought II》很强,但与《Deep Thought》相比,它的玩法似乎缺乏活力。在马克西姆·德鲁吉看来,“深思二号”下棋很扎实,但它似乎没有能力像“深思”那样下出令马克西姆惊讶的棋步。我不能断定托马斯的论文是错误的,但实验室测试似乎表明有理由怀疑。也许进步比托马斯的结论还要大。

Deep Thought II had up to twenty-four chess processors, and the new chess program was designed to use all of them. The first version of the program did not implement any of the selective extensions used in Deep Thought, nor did it implement the search ideas that I had. It took several months to complete. The new program was about 5–8 times faster than Deep Thought. In his thesis, Thomas claimed that the search extensions used in Deep Thought produced only a small improvement in performance, which would not have been enough to compensate for the 5–8 times advantage in speed that Deep Thought II enjoyed. Yet, Deep Thought was on even footing with the first version of Deep Thought II in head-to-head competition. Maxim Dlugy had already left for Bankers Trust by then, but he still had some IBM contract hours to fill and he played a game against Deep Thought II. After the game, he opined that Deep Thought II was strong, but its play seemed to lack pizazz when compared with Deep Thought. In Maxim Dlugy’s eyes, Deep Thought II played solid chess, but it did not seem to be capable of playing moves that surprised Maxim as Deep Thought had. I could not say conclusively that Thomas was wrong in his thesis, but the lab tests seemed to indicate there were reasons to doubt. Maybe the improvement was bigger than Thomas had concluded.

新的选择性搜索想法又花了几个月的时间才得以实施。起初,Deep Thought 仍然凭借 Deep Thought II 保持着自己的地位。我们为 Deep 安排了一些公共活动Thought II,但由于 Deep Thought 看起来与 Deep Thought II 一样好(或更好),而且也是更稳定的程序,因此我们选择让 Deep Thought 在事件中发挥作用。然而,组织者已经公开宣称这些活动以 Deep Thought II 为主题。正如前面所述,在其中一次活动中,我们发现需要对 Maxim Dlugy 的开篇书籍进行进一步的研究。那时,他的合同已经到期。穆雷最终放弃了这本书,并编写了一个程序,利用从公开比赛记录中收集到的顶级球员的开局曲目,自动生成新的开局曲目。他创作的一本开场书籍是根据阿纳托利·卡尔波夫的开场曲目改编的。这是通过从阿纳托利的所有公开游戏中提取最常见的动作来实现的11。这本开篇书随后被用于 Deep Thought II 的首次公开亮相。

The new selective search ideas took a few more months to implement. At first, Deep Thought was still holding its own with Deep Thought II. We had a few public events scheduled for Deep Thought II but since Deep Thought seemed as good as (or better than) Deep Thought II and was also the more stable program, we opted to let Deep Thought play in the events. The organizers, however, had already publicly billed the events as featuring Deep Thought II. As recounted earlier, during one of these events, we discovered the need of further work on Maxim Dlugy’s opening book. By then, his contract had expired. Murray eventually gave up on the book and wrote a program to automatically generate new opening books, using the top players’ opening repertoires gleaned from the public game records. One of the opening books that he generated was based on Anatoly Karpov’s opening repertoire. This was accomplished by extracting the most common moves from all of Anatoly’s public games11. This opening book was subsequently used in Deep Thought II’s public debut.

有一段时间,我所有的软件工作似乎都白费了。尽管拥有额外的处理能力,Deep Thought II 并不比 Deep Thought 更好。然而,事实证明,实践经验是无价的,我当然更充分地欣赏托马斯所做的事情。我还对如何设计硬件以简化软件工作有了更好的想法。但更重要的是,我对如何更有效地控制搜索有了一些新的见解。正是有了这些新的见解,才完成了最后的润色。结果是戏剧性的。Deep Thought II 突然超越了 Deep Thought。Deep Thought 已经是一个国际象棋战术怪物,然而,在一场又一场的比赛中,Deep Thought II 发现了深度组合,通过一些额外的动作逃脱了 Deep Thought。我们有一个新的、排名第一的国际象棋程序。

For a while, it appeared that all my software work was for naught. Deep Thought II was no better than Deep Thought, despite all the extra processing power. The hands-on experience, however, proved invaluable and I certainly got to appreciate what Thomas did more fully. I also had a better idea about how to design the hardware to ease the software work. But, more importantly, I got some new insights into how to control the search more effectively. It was with these new insights that the final touch was added. The result was dramatic. Deep Thought II was suddenly outclassing Deep Thought. Deep Thought was already a chess tactical monster and yet, in game after game, Deep Thought II found deep combinations that escaped Deep Thought by a few extra moves. We had a new, number one chess program.

Deep Thought II 于 11 月 17 日至 20 日在新墨西哥州阿尔伯克基举行的 1991 年 ACM 计算机国际象棋锦标赛上进行了首次公开比赛。没有使用德鲁吉开局书,而是使用了“卡尔波夫开局书”。Deep Thought II 赢得了所有比赛;然而,在其中一场黑棋中,我们感到害怕,因为卡尔波夫的书允许我们的对手在开局时选择重复抽牌——众所周知,阿纳托利愿意以黑牌抽牌。对我来说,Deep Thought II 在本次比赛中最令人印象深刻的比赛是对阵 Cray Blitz。对于超过十步(白方和黑方各十步,总共二十层),Cray Blitz 完全没有意识到自己会被杀死,而 Deep Thought II 则预测整个游戏的继续,并评估游戏完全获胜。Cray Blitz 是最快的国际象棋程序之一,其搜索速度如此之快超出了我最疯狂的预期。

Deep Thought II played its first public event at the 1991 ACM Computer Chess Championship, held in Albuquerque, New Mexico from November 17–20. Instead of using the Dlugy opening book, the “Karpov opening book” was used. Deep Thought II won all its games; however, in one of its Black games, we had a scare, as the Karpov book allowed our opponent the option of getting a repetition draw right at the opening—Anatoly was well known to be willing to draw as Black. To me, the most impressive game played by Deep Thought II in this event was against Cray Blitz. For over ten moves (ten moves each by White and Black, or twenty plies total), Cray Blitz had no inkling that it was getting killed, while Deep Thought II was predicting the entire game continuation and assessing the game as completely won. Cray Blitz was one of the fastest chess programs around, and to outsearch it by this wide margin exceeded my wildest expectation.

阿尔伯克基比赛结束后,乔·霍恩 (Joe Hoane) 接管了国际象棋项目。当我交付时,Deep Thought II 只有大约 5000 行代码,大约是 Deep Thought 代码行数的二十分之一。Deep Thought II 代码处理的国际象棋处理器数量是原来的十倍,而且它还具有新的、可能更强大的搜索扩展。我为这项工作感到自豪,但软件工作花费了我大约一年的时间,否则这些时间本可以花在国际象棋芯片的设计上。

Joe Hoane took over the chess program after the Albuquerque event. At the time I handed it over, Deep Thought II had only about 5000 lines of code, or roughly one-twentieth the number of lines of code in Deep Thought. The Deep Thought II code handled ten times more chess processors, and it also had new and possibly far more powerful search extensions. I was proud of the work, but the software effort took about a year of my time which would otherwise have been spent on the chess chip design.

丹麦之旅

A Trip to Denmark

多年来,在 ChipTest、Deep Thought 或 Deep Thought II 参加的每场大型赛事之后,我都会在 USENET 新闻组rec.games.chess中发布我对赛事的个人帐户12。以下关于我们丹麦之行的报告改编自其中一篇文章。Deep Thought II 的职业生涯中还有其他重要的国际象棋赛事,包括战胜大师朱迪特·波尔加 (Judit Polgar),她是世界顶级棋手之一,也可能是有史以来最好的女棋手13。然而,由于许多其他原因,丹麦的国际象棋赛事也很重要。历史上,这是国际象棋计算机首次在一场比赛中击败国家队。从深蓝战队的角度来看,这也是我们第一次参加正式的古典象棋比赛(国际象棋在锦标赛时间控制下进行),从这个意义上说,这对深蓝队的比赛来说是一次预演。来。这次比赛,特别是与狡猾的本特·拉森的比赛,为我们提供了许多有用的国际象棋课程。我把这一事件写进书中还有一个非常重要的原因。乔·霍恩 (Joe Hoane) 接管了节目制作事件发生时,《Deep Thought II》刚刚上线一年多。此时,他已经对该程序进行了重大的补充和改进,Deep Thought II 或多或少已经成为了他的程序。但直到这次活动,我才第一次真正看到乔对这个项目的投入。乔在活动期间生病了,但他坚持了下来,长时间工作,并对程序进行了许多改进。在这次丹麦之行中,穆雷、乔和我真正成为了一个团队。

Over the years, after every big event that ChipTest, Deep Thought, or Deep Thought II played in, I would post in the USENET news-group rec.games.chess my personal account of the event12. The following report on our trip to Denmark was adapted from one of the posts. There were other important chess events in the career of Deep Thought II, including a match win over Grandmaster Judit Polgar, one of the top players in the world and possibly the best female player ever13. The chess event in Denmark, however, was important for many other reasons. Historically, it was the first time that a chess computer defeated a National Team in a match. From the Deep Blue team’s perspective, it was also the first time that we played in a serious match of classical chess (chess played at tournament time control), and in that sense, it was a dry run for the Deep Blue matches that were to come. The event, especially the games against the wily Bent Larsen, provided many useful chess lessons for us. There is one more very important reason why I included this event in the book. Joe Hoane had taken over the programming of Deep Thought II for just over a year when the event took place. By this time, he had made significant additions and improvements to the program, and Deep Thought II had more or less become his program. But it was only in this event that I really got to see Joe’s dedication to the project for the first time. Joe was sick during part of the event, but he hung in there throughout, worked long hours, and made numerous improvements to the program. Murray, Joe, and I really solidified as a team during this trip to Denmark.


阿尔伯克基事件之后我能够重新回到芯片设计,但发现早期的设计工作不再有用。休息时间太长了,我忘记了自己在做什么。与此同时,我们的经理兰迪·莫利克 (Randy Moulic) 有其他想法。IBM 正在与另一家公司组建一家合资企业,兰迪也将参与其中。Chung-Jen (CJ) Tan 于 1992 年春天被任命为我们的新经理,但由于 Murray、Joe 和我都因自己的原因参与了该项目,并且实际上是自主的,因此管理链的变化对我们的影响很小。我们的日常生活。对我来说,这个项目就是我的生活,对于默里和乔来说当然也很相似。无论经理是谁,我们都会看到项目完成。就我们而言,经理只是一个辅助角色,即使我们从呆伯特漫画中找到尖头发的老板,我们也能完成工作14。CJ 发挥了很好的支持作用,因为团队基本上不知道他或官僚机构在做什么。该团队很好地免受外界干扰。

I was able to get back to the chip design after the Albuquerque event, but found that the earlier design work was no longer useful. It was too long a break and I lost track of what I had been doing. Meanwhile, our manager, Randy Moulic, had something else on his mind. IBM was forming a joint venture with another company, and Randy was going to be involved. Chung-Jen (C J) Tan was assigned as our new manager in the spring of 1992 but since Murray, Joe and I were all in the project for our own reasons, and effectively autonomous, the change in the management chain had very little impact on our day-to-day routine. To me, the project was my life and certainly close to the same thing for Murray and Joe. We would see the project completed no matter who the manager was. As far as we were concerned, the manager was a support role and, even if we were to get the pointy-haired boss from the Dilbert comic strip, we would have gotten the job done14. C J played a fine support role as the team was largely unaware of what he or the bureaucracy was doing. The team was well insulated from outside disturbances.

当我重新开始我的芯片设计工作时,穆雷发现了互联网国际象棋服务器,而深思二号在被发现之前在服务器上匿名玩了一段时间。Deep Thought II 在此期间调试得相当好,为我们的下一次活动做好了良好的准备。

While I was restarting my chip design work, Murray discovered the Internet Chess Server, and Deep Thought II played anonymously on the server for a while before it was discovered. Deep Thought II became fairly well debugged during this period, and it was good preparation for our next event.

1992 年末,IBM 丹麦公司与我们讨论安排国际象棋表演赛,可能与 IBM 瑞典公司有关。在某个时候,有人谈论玩瑞典语特级大师赛也是如此,但最终的安排是只在丹麦哥本哈根比赛,并且只对阵丹麦选手。

In late 1992, IBM Denmark talked with us about arranging chess exhibition matches, possibly in connection with IBM Sweden. At some point, there was talk about playing Swedish Grandmasters as well, but the final arrangement was to play only in Copenhagen, Denmark, and only against Danish players.

比赛安排了两场单独的正式比赛,一场是对阵总经理本特·拉森的四场比赛,一场是对阵丹麦国家队的四场比赛,最后一场对阵本特的比赛也算作团体赛的一部分。所以,实际上只有七场正式比赛。工作日还安排了另外五场一小时的表演赛。哥本哈根国际象棋联盟与丹麦 IBM 共同赞助了本次比赛。为了这次活动,我们亲自将国际象棋处理器的印刷电路板运到丹麦,并将 Deep Thought II 的其余部分装在一个大板条箱中运送。

Two separate official matches were scheduled, a four-game match against GM Bent Larsen and a four-game match against the Danish national team, with the last game against Bent also counted as part of the team match. So, there were actually only seven official games. Five additional one-hour exhibition games were also scheduled during the weekdays. The Copenhagen Chess Union co-sponsored the matches together with IBM Denmark. For this event, we hand carried the printed circuit boards for the chess processors to Denmark and shipped the rest of Deep Thought II in a large crate.

对阵丹麦选手的主要比赛将于 1993 年 2 月 24 日星期三开始。我们于 2 月 19 日星期五提前到达,以便参加第二天与加里·卡斯帕罗夫一起举行的公开比赛。抵达当晚,我们与卡斯帕罗夫和组织者共进晚餐。晚餐时,我们首先了解到丹麦人准备得有多充分。Jens Nielsen 是著名的计算机国际象棋棋局测试集的创建者,他为丹麦棋手制作了一本笔记本,其中包含 Deep Thought 游戏、相关文章、有关计算机弱点的提示等等。我们很担心,但 Deep Thought II 已经支离破碎,并且有预先安排的周末活动。周一之前我们无能为力。

The main event against the Danish players was to start on Wednesday 24 February 1993. We arrived early on Friday 19 February in order to take part in a public event the next day with Garry Kasparov. We had dinner with Kasparov and the organizers in the evening of our arrival. During dinner, we first learned how well prepared the Danes were. Jens Nielsen, the creator of a well known test set of chess positions for computers, had made a notebook for the Danish players, containing the Deep Thought games, relevant articles, pointers about computers’ weaknesses, and so on. We were concerned, but Deep Thought II was in pieces, and there were prescheduled weekend events. There was not much we could do before Monday.

加里主导了晚餐的谈话,但他时不时地会退缩起来,只是倾听。也许是他感冒了。但也有可能他心里有别的事情。几天后,在我们的一场比赛中,我们收到一份传真,称加里·卡斯帕罗夫(Garry Kasparov)和他的英国挑战者奈杰尔·肖特(Nigel Short)决定离开国际象棋联合会(FIDE)。他们正在组建一个新的国际象棋组织——职业国际象棋协会(PCA),并将在 PCA 的主持下进行世界冠军赛。奈杰尔不喜欢国际棋联在未经他同意的情况下对世界锦标赛比赛强加的条件,加里与时任国际棋联主席弗洛伦西奥·坎波马内斯之间存在长期争执,奈杰尔找到了一个愿意的合作伙伴。

Garry dominated the dinner conversation, but every once in a while he withdrew into himself and just listened. Maybe it was his cold. But it could also have been the case that something else was on his mind. A few days later, during one of our match games, we received a fax that Garry Kasparov and Nigel Short, his British Challenger, had decided to leave FIDE, the international chess federation. They were forming a new chess organization, Professional Chess Association (PCA), and would play their World Championship match under the auspices of PCA. Nigel did not like the conditions FIDE was imposing on the World Championship match without his consent, and with Garry in a prolonged dispute with the then FIDE president Florencio Campomanes, Nigel found a willing partner.

尽管感冒了,加里在周六的公开活动中仍然给人留下了深刻的印象。我们无法为该节目制作完整的《Deep Thought II》,但我们设法获得了一个只有两个国际象棋处理器启动并运行的版本。加里走过其中之一他最近对阵总经理普雷德拉格·尼科利奇的比赛。在这场比赛中,加里放弃了主要的进攻材料,在他看来,这显然是胜利。Deep Thought II的处理能力有所下降,最初评估该位置对Garry不利,并为对方提供了一些替代防御。随后,Garry 在舞台上实时打破了 Deep Thought II 给出的新防御。Deep Thought II显然对游戏位置上发生的事情一无所知。我并没有完全感到沮丧和厄运,但很明显,当我回到美国时,我将有很多工作要做。所涉及的问题或多或少是我一个人的问题——新的国际象棋芯片必须“理解”这场比赛中的局面15

Garry was impressive on Saturday during the public event, despite having a cold. We could not put together the full Deep Thought II for the show, but we managed to get a version with only two chess processors up and running. Garry went over one of his recent games against GM Predrag Nikolic. In this game, Garry gave up major material for an attack, that in his judgment, was clearly winning. Deep Thought II, with reduced processing power, assessed the position as unfavorable for Garry initially, and gave some alternative defenses for the other side. Garry then proceeded to bust the new defenses given by Deep Thought II in real time on the stage. Deep Thought II clearly did not have a clue about what was going on in the game position. I was not precisely feeling gloom and doom, but it was clear that I would have a lot of work to do when I got back to the United States. The problems involved were more or less mine alone—the new chess chip would have to “understand” the positions in this game15.

我们于 2 月 22 日星期一开始在比赛现场安装完整的机器。机房没有充分冷却(没有空调),其中一台国际象棋处理器过热并在几个小时后停止运行。周二,通风状况得到改善,并进行了一场测试比赛,没有发生进一步的事故。

We started setting up the full machine at the match site on Monday 22 February. The machine room was not adequately cooled (there was no air conditioning) and one of the chess processors overheated and stopped functioning after a few hours. The ventilation was improved by Tuesday and a test game was played without further mishap.

该机器被组织者称为“北欧深蓝”。这是“深蓝”这个名字第一次被使用,尽管是间接的。就我而言,深蓝直到1996年完成第一场对阵加里的比赛后才存在。丹麦实际运行的机器是 Deep Thought II,尽管国际象棋处理器的数量有所减少。在 Deep Thought II 首次公开亮相时,它使用了 24 个国际象棋处理器,但此时我们只使用了 14 个处理器。除了一个处理器因过热而丢失外,其他处理器也在我们实验室的一场怪异事故中丢失。我们实验室的空调结冰并停止工作。然后冰融化,滴在国际象棋处理器上,导致其中一些处理器出现故障。在哥本哈根赛事中,Deep Thought II 的峰值速度降至每秒约 4-5 百万个国际象棋棋局。

The machine was referred to as Nordic Deep Blue by the organizers. This was the first time that the Deep Blue name was used, albeit obliquely. As far as I am concerned, Deep Blue did not exist until it was completed in 1996 for its first match against Garry. The machine that actually played in Denmark was Deep Thought II, although with a reduced number of chess processors. In Deep Thought II’s public debut it played with twenty-four chess processors but at this time we were using only fourteen processors. Besides the one processor lost from the overheating, others were lost in a freak accident in our lab. The air conditioner in our lab iced up and stopped working. The ice then melted, dripped on the chess processors and caused some of them to malfunction. The peak speed of Deep Thought II was down to about 4–5 million chess positions per second for the Copenhagen events.

第一天(2 月 24 日星期三)

Day 1 (Wednesday, 24 February)

这是灾难性的一天。我们第一次充分认识到自己准备不足的严重性以及准备不足的程度。我们的反对派已经做好了准备。比赛结束后,紧接着就是紧张的工作。

This was a disastrous day. It was the first time that we fully realized the seriousness of our own lack of preparation and how well prepared was our opposition. Frantic work followed after the games.

第一场比赛对阵本特·拉森。这是对阵他的第一场正式比赛,而不是对阵丹麦队的比赛的一部分。本特采取了一个简单的策略,效果出奇的好。当国际象棋棋局开阔且许多“开放”的对角线不受棋子限制时,在对角线上操作的象会更加强大。如果棋位被兵挡住,马可能会变得比象更强大,尽管象通常是上等棋子。本特用掉了深思二号的所有骑士,试图封锁阵地。Deep Thought II保留了一对主教,这通常被认为是一个优势。问题是我们从来没有对 Deep Thought II 进行编程来理解它应该交换一些棋子来增加象对的范围。如果我们事先预见到这是一个潜在的问题,本特的策略就不会奏效。我们的开幕准备工作也不尽如人意。本特赢得了第一场比赛,而我们却陷入了深渊。

The first game was against Bent Larsen. This was the first official game of the match against him, and not part of the match against the Danish team. Bent adopted a simple strategy that worked surprisingly well. Bishops, operating on diagonals, are more powerful when the chess position is wide open with many “open” diagonals uninhibited by pawns. If the chess position is blocked by pawns, the knights might become more powerful than the bishops, even though bishops are normally superior pieces. Bent traded off all the knights belonging to Deep Thought II and tried to keep the position blocked. Deep Thought II retained a pair of bishops, which is usually considered an advantage. The problem was that we had never programed Deep Thought II to understand that it should trade some pawns to increase the scope of the bishop pair. Bent’s strategy would not have worked if we had forseen this as a potential problem beforehand. Our opening preparation also left something to be desired. Bent won the first match game, and we were in a deep hole.

正式比赛后我们进行了两场简短的表演赛。两人再次以灾难告终。我们的对手早有准备,深思二号似乎出了问题。

We played two short exhibition games after the official game. Again both ended in disaster. Our opponents were well prepared, and there seemed to be something wrong with Deep Thought II.

第 2 天(2 月 25 日,星期四)

Day 2 (Thursday, 25 February)

前一天晚上,对程序进行了一些修改。乔通宵达旦。穆雷也工作到很晚。我主要是一个旁观者。除了 Deep Thought II 在阿尔伯克基首次亮相之外,当我编写大部分国际象棋程序时,我的主要工作通常会在国际象棋比赛时完成。一夜之间,Joe 发现 Deep Thought II 的搜索扩展代码存在一些问题。在我们访问丹麦之前,Joe在程序中放入了一些调试代码进行测试,并且这些代码在第一天的比赛中是活跃的。调试代码导致程序的行为完全不同,而且看起来要弱得多。乔找到了问题并纠正了它们。

The night before, a few modifications were made to the program. Joe put in an all-nighter. Murray also worked late. I was mostly a spectator. With the exception of the debut of Deep Thought II in Albuquerque, when I wrote the bulk of the chess program, my main work would have been completed usually by the time of a chess event. Overnight, Joe found some problems in the search extension code of Deep Thought II. Prior to our visit to Denmark, Joe put some debugging code into the program for testing, and this code was active during the games on the first day. The debugging code caused the program to behave quite differently and seemingly much weaker. Joe located the problems and corrected them.

白天进行了三场表演赛,Deep Thought II 的表现比前一天好很多。我们此次哥本哈根之行的日程非常紧凑,加上表演赛的不断进行,我们没有太多时间认真准备正式比赛。丹麦队可能利用表演赛来寻找 Deep Thought II 的其他弱点。

Three exhibition games were played during the day, and Deep Thought II behaved much better than it had the previous day. The schedule for our Copenhagen visit was exceedingly tight, and with the continuous exhibition games, we did not have much time to make serious preparation for the official matches. The Danish team was probably using the exhibition games to look for additional weaknesses in Deep Thought II.

第二场比赛,也是团体赛的第一场比赛,在晚上对阵丹麦国家队的 IM Henrik Danielsen。Henrik 对 Deep Thought II 的几个犀利动作感到惊讶。乔深夜的改变似乎产生了很大的变化。在 Deep Thought II 做出了一个特别令人不快的举动后,Henrik 花了十五分钟多的时间思考他的反应。机器正确地预测了他的反应,并能够利用他在该位置上花费的时间来想出一个几乎结束了比赛的杀手锏。如果亨里克反应快的话,“深思二号”就会错过这一杀手锏。

The second match game, and first game of the team match, was played in the evening against IM Henrik Danielsen of the Danish national team. Henrik was surprised by several sharp moves from Deep Thought II. Joe’s late night changes seemed to have made a big difference. After one particularly unpleasant move from Deep Thought II, Henrik spent over fifteen minutes thinking about his response. The machine correctly predicted his response and was able to use the time he spent on the position to come up with a killer move that pretty much ended the game. If Henrik had responded quickly, Deep Thought II would have missed the killer move.

于是两天后,Deep Thought II在与拉森队的比赛中0-1,在对阵丹麦队的比赛中1-0。那天晚上我们睡得更安稳了。

So after two days, Deep Thought II was 0–1 in the Larsen match, and 1–0 in the match against the Danish team. We slept easier that night.

第 3 天(2 月 26 日星期五)

Day 3 (Friday, 26 February)

这是安静的一天。对阵本特·拉森和丹麦队的两场比赛均以平局告终。

It was a quiet day. Both match games against Bent Larsen and the Danish team, ended in draws.

Deep Thought II 在对阵 Bent 的第二场比赛中有很大机会获胜,但未能成功。整场比赛他都处于压力之下,比赛进展到双方各有一车一象,深思二号多了一个棋子的结局。问题是主教不在相同颜色的方格上。众所周知,双方在相反颜色的方格上都有一个主教的结局对于较强的一方来说是很难获胜的,因为每个主教只能攻击棋盘的一半。深思二号知道这一点;不幸的是,它不知道双方只增加一辆车并不能改变局面。

Deep Thought II had good chances to win the second match game against Bent, but could not pull it off. He was under pressure throughout the game, and the game progressed into an ending where both sides had a rook and a bishop and Deep Thought II had an extra pawn. The problem was that the bishops were not on squares of the same color. It is well known that endings where the two sides have a bishop on the opposite color squares are notoriously difficult to win for the stronger side, since each bishop can only attack half of the chess board. Deep Thought II knew about this; unfortunately, it did not know that having only one additional rook on both sides did not change the situation much.

当天的第二场比赛对阵丹麦队成员IM Carsten Hoi。卡斯滕取得了良好的开局位置,部分原因是我们开局准备工作中的疏忽。鉴于处于弱势,“深思二”发挥强劲,给卡斯滕带来了棘手的选择。在一个位置上,“深思二号”似乎为他提供了一个免费的棋子。卡斯滕想了很长时间才拒绝了典当提议,比赛以平局结束。几乎所有观众都希望卡斯滕能拿下棋子并赢得比赛。他事后解释说,他并没有看出拿走棋子有什么明显的问题,但他的直觉告诉他,这可能是一个陷阱。Deep Thought II 的赛后分析表明卡斯滕的直觉非常好。如果他拿走了棋子,要么他的骑士被困住,要么“深思二号”会受到非常强烈的攻击。

The second game of the day was against the Danish team member IM Carsten Hoi. Carsten achieved a good opening position in part owing to an oversight in our opening preparation. Deep Thought II played strongly, given the weak position, and presented Carsten with tricky choices. In one position, Deep Thought II appeared to be offering him a free pawn. Carsten thought for a very long time before declining the pawn offer and the game ended in a draw. Almost everyone in the audience expected Carsten to take the pawn and win the game. He explained afterwards that he did not see anything obviously wrong with taking the pawn, but his intuition told him that it might be a trap. Postgame analysis with Deep Thought II suggested that Carsten’s intuition was excellent. Had he taken the pawn, either his knight would have been trapped or Deep Thought II would have had a very strong attack.

第 4 天(2 月 27 日星期六)

Day 4 (Saturday, 27 February)

本特·拉森再次拿下了怀特。穆雷在第一场比赛中找不到本特战术的良药,所以老备用中卫再次抬头。当然,在丹麦,它更广为人知的名字是斯堪的纳维亚防御,而《Deep Thought II》恰好扮演了“丹麦变体”。我们防守并不是为了向东道主致敬。我们很绝望。比赛开始后不久,深思二号就遇到了大麻烦。比赛现场中唯一不“认为”本特获胜的下棋实体可能是 Deep Thought II。但即使是机器也认为自己的处境非常不利。然而,本特确实有一个心理问题。进入比赛后,他可能期待着另一次缓慢的位置挤压。棋盘上的位置对他来说非常有利,但他必须完美地执行进攻。如果本特面对的是人类玩家,那么人类对手可能会在压力下崩溃。如果他面对的是另一台计算机,那么计算机很有可能会犯一些严重错误并着火。本特的进攻一度表现出色,但“深思二号”却走得很巧妙,防守也很机智。慢慢地,他的优势就消失了。当进入残局阶段时,Deep Thought II 占据了更好的位置。关于最终位置是否可以获胜存在一些争论,但由于 Deep Thought II 不知道正确的前进方式,我们宣布休战。

Bent Larsen had White again. Murray could not find a good antidote to Bent’s tactics in the first game so the old standby, the Center Counter Defense, reared its head again. Of course, in Denmark, it is better known as the Scandinavian Defense, and Deep Thought II happened to play the “Danish variation”. We did not play the defense to honor our host. We were desperate. Shortly after the start of the game, Deep Thought II was in big trouble. Probably the only chess-playing entity at the match site that did not “think” that Bent was winning was Deep Thought II. But even the machine assessed its position as highly unfavorable. Bent, however, did have one psychological problem. Going into the game, he was probably expecting another slow positional squeeze. The position over the board was very good for him, but he would have to execute the attack perfectly. If Bent were facing a human player, the human opponent probably would have cracked under the pressure. If he were facing another computer, there would have been a very good chance that the computer would have made some critical mistake and gone down in flames. Bent executed the attack well for a while, but Deep Thought II walked a fine line and defended resourcefully. Slowly, his advantage dissipated. When the endgame phase was reached, it was Deep Thought II that had the better position. There were some debates as to whether the final position was winnable, but since Deep Thought II had no idea about the proper way to proceed, we called a truce.

对阵丹麦国家队第三位队员总经理拉尔斯·博·汉森的团体赛也许是 Deep Thought II 在丹麦打得最好的一场比赛。一位评论员甚至说《深思二》的玩法就像一位超级大师。拉尔斯此前曾在对阵 Deep Thought II 的开场线时取得过成功,但赛后表示他可能不得不重新考虑整条线。与 Lars 的比赛是我们第一次看到 Deep Thought II 在常规锦标赛中以优越的位置发挥战胜了 Grandmaster 。最后的致命一击是战术组合,但比赛在此之前就已经结束了。本场比赛结束,Deep Thought II以2.5的比分领先丹麦国家队,锁定胜局。0.5。我们仍然以 2 比 1 落后本特,并且我们还有最后一枪扳平了拉森队的比赛。

The team match game against the third Danish National Team member, GM Lars Bo Hansen, was perhaps the best game played by Deep Thought II in Denmark. One commentator went so far as to say that Deep Thought II played like a Super Grandmaster. Lars had had previous successes against the opening line played by Deep Thought II, but said after the game that he might have to rethink the whole line. The game against Lars was the first time that we saw Deep Thought II outplay a Grandmaster in a regular tournament game by superior positional play. The final coup de grace was a tactical combination, but the game was finished before that. After this game, Deep Thought II clinched the match against the Danish National Team, leading by the score of 2.5 to 0.5. We were still behind Bent by 2 to 1, and we had one last shot to tie the Larsen match.

第 5 天(2 月 28 日星期日)

Day 5 (Sunday, 28 February)

拉森队比赛和丹麦国家队比赛的主播比赛都在这一天进行。本特很早就陷入了麻烦,低估了关键的一步。Deep Thought II 可能在第 25 手棋时具有获胜优势,但随后它决定以相反颜色的主教进入另一个结局。这一次,双方都各有一位女王。第二局结束后,我们对双方都多了车的情况加大了处罚力度,但对双方都有后的情况没有这样做。我们犯傻了吗?Deep Thought II 未能取得任何进展。过了一会儿,我们开始在机房里讨论是否应该给本特平局。IM Bjarke Kristensen 非常兴奋地走进机房,并开始向我们解释,如果皇后被交易的话,这个位置对于 Deep Thought II 来说是胜利!听到这个消息我们很震惊。穆雷是队里的国际象棋专家,他不知道这种可能性。比亚克告诉我们,一旦交换了皇后,就会赢得纯粹的“相反颜色的主教”结局。Deep Thought II 然后可以创建两个相距较远的“通过棋子”(棋子不再被敌方敌方棋子抑制或阻挡),尽管主教颜色相反,但它们仍会获胜。然而,我们遇到的问题是,Deep Thought II 被明确编程为在这种情况下不会牺牲其女王。我们又玩了一会儿,然后向本特提出了平局。比赛结束后,我对这种经历感到彻底厌恶,我翻阅了有关国际象棋结局的书籍,并将所有“简单”残局规则的知识,包括“相反颜色的主教”结局的详细规则,添加到新的国际象棋筹码中。

The anchor game for both the Larsen match and the Danish National Team match was played on this day. Bent got into trouble early, underestimating a critical move. Deep Thought II probably had a winning edge at move twenty-five, but then it decided to go into another ending with bishops of opposite color. This time, both sides had a queen as well. After the second game, we increased the penalty for having an additional rook for both sides, but did not do so for having a queen for both sides. Did we goof? Deep Thought II was not able to make any progress. After a short while, we started discussing in the machine room whether we should just offer Bent a draw. IM Bjarke Kristensen came into the machine room very excited and started to explain to us that the position was a win for Deep Thought II, if the queens were traded off! We were shocked to hear that. Murray was the chess expert of the team, and he did not know about the possibility. Bjarke told us that once the queens were traded off, the pure “bishop of opposite color” ending would be won. Deep Thought II could then create two widely separated “passed pawns” (pawns that can no longer be inhibited or blocked by opposing enemy pawns), which would win despite the bishops of opposite color. The problem we had, however, was that Deep Thought II was explicitly programmed not to trade off its queen in this situation. We played on a little while longer, and then offered Bent the draw. After the match, thoroughly disgusted with this experience, I went through books on chess endings and added knowledge about all the “easy” endgame rules, including detailed rules on the “bishop of opposite color” endings, to the new chess chip.

Larsen 比赛的最终得分为 Larsen 2.5,Deep Thought II 1.5。对阵丹麦国家队的结果要好得多:Deep Thought II 3和丹麦队1。今天回顾,结果已经不再重要,但比赛经验和我们从这次旅行中获得的国际象棋课程是无价的。

The final tally for the Larsen match was Larsen 2.5 and Deep Thought II 1.5. The result was much better against the Danish National Team: Deep Thought II 3 and Danish team 1. The results, looking back today, are no longer important, but the match experience and the chess lessons we obtained from the trip were invaluable.

    如今,IBM 研究院的文书工作量要少得多。计算环境也更接近你在外界看到的情况。

    The amount of paperwork is much lower in IBM Research these days. The computing environment is also closer to what you see in the outside world.

    Deep Blue 是对 IBM 昵称 Big Blue 的演绎,该昵称可能来自其商标的蓝色。IBM 不拥有“蓝色巨人”的商标,也从未使用该术语来描述自己。

    Deep Blue is a play on IBM’s nickname Big Blue, which probably came from the blue color of its trademark. IBM does not own the trademark for “Big Blue” and never uses the term to describe itself.

    几年后,彼得·布朗在一位强大高手的帮助下,与一个非常早期版本的深蓝小号进行了一场休闲比赛,结果输了。彼得想成为第一个“玩它”的人。

    A few years later, Peter Brown, with the help of a strong master, played a casual game against a very early version of Deep Blue Jr. and lost. Peter wanted to be the first person to “play it”.

    然而,商业国际象棋程序员总是将 ChipTest 和 Deep Thought 称为“超级计算机”,暗示它们只是输给了价值数百万美元的机器。

    The commercial chess programmers, however, always referred to both ChipTest and Deep Thought as “supercomputers”, implying that they only lost to multi-million-dollar machines.

    这发生在 Deep Thought II 首次部分运行时。另请参阅《深思熟虑 II》和《更绿的牧场》的故事。据我的消息来源称,除了马克西姆·德鲁吉(Maxim Dlugy)之外,信孚银行还聘请了另外两名国际象棋棋手。

    This happened at about the time when Deep Thought II first became partially operational. See also the story on Deep Thought II and the Greener Pasture. According to my sources, Bankers Trust hired two other chess players in addition to Maxim Dlugy.

    这场比赛被宣传为 Deep Thought II 比赛,但 Deep Thought 才是选手。

    The match was billed as a Deep Thought II match, but Deep Thought was the player.

    不要问我程序做了什么。托马斯从未详细解释过。

    Don’t ask me what the program did. Thomas never explained it in detail.

    Software Toolworks 锦标赛结束后,IM Mike Valvo 向 Deep Thought 发起了一场两局通信赛的挑战,他轻松获胜。信件游戏强烈表明,让程序沿着“有趣”的路线进行更深入的搜索可能是个好主意。

    After the Software Toolworks Championship, IM Mike Valvo challenged Deep Thought to a two-game correspondence match that he won easily. The correspondence games strongly suggested that it might be a good idea to have the program search a great deal deeper along “interesting” lines.

    Murray 是一位不错的程序员,但当时只有 Thomas 和我可以编写新程序,因为我们是仅有的两个对硬件、并行搜索和选择性搜索有深入了解的人。

    Murray was a decent programmer, but at the time, only Thomas and I could possibly write the new program as we were the only two with intimate knowledge of the hardware, the parallel search, and the selective search.

   人数是一个招聘职位。

   A head count is a hiring position.

   为什么穆雷选择使用阿纳托利的开场曲目而不是加里的?嗯,阿纳托利的开局曲目相对狭窄且安全,而加里的开局曲目非常广泛,变化迅速,有时甚至有风险。要确保“卡斯帕罗夫开局书”可以安全使用,需要付出巨大的努力。

   Why did Murray choose to use Anatoly’s opening repertoire instead of Garry’s? Well, Anatoly played with a relatively narrow and safe opening repertoire while Garry’s was very wide, rapidly changing, and sometimes risky. It would have taken a great deal of effort to ensure that a “Kasparov opening book” was safe to use.

   在第一场“深蓝”比赛之后,我写了一个个人帐户,但当 IBM 研究部通信部门说服我最好将其作为文章发表在一本严肃的杂志上时,我决定不发布它。由于第二场比赛的准备工作正在进行中,这篇文章从未见诸世人。第二场深蓝比赛,我没有写任何赛后报告。

   After the first Deep Blue match, I wrote a personal account but decided against posting it when the Communications Department at IBM Research persuaded me that it was probably better to publish it as an article in a serious magazine. As preparations for the second match were under way, the article never saw the light of the day. For the second Deep Blue match, I did not bother to write any post-match report.

   这场比赛是在比赛时间控制下进行的(整场比赛每名球员30分钟),不应该太认真。

   The match was played under action time control (30 minutes per player for the entire game) and should not be taken too seriously.

   好吧,我承认,如果那个尖头发的老板是​​我们的经理,那就很难了。很可能,这位尖头发的老板会宣称自己是我们无所畏惧的项目负责人,并命令我们在他强迫我们参加的所有毫无意义的会议上为他提供免费的甜甜圈。

   Okay, I will admit that it would have been very hard if the pointy-haired boss had been our manager. In all likelihood, the pointy-haired boss would have proclaimed himself our fearless project leader and ordered us to provide him free donuts at all the meaningless meetings he forced upon us.

   如今,计算能力仅为深蓝三十分之一的 Deep Blue Jr. 在大约一分钟内播放牺牲本身没有任何问题。所以那天我确实从加里那里学到了一些有用的东西。国际象棋芯片需要认识到某些棋子配置,与对抗敌方国王的开放线路相结合,在国王狩猎中可能非常强大。

   Today, Deep Blue Jr., which has only one-thirtieth the computation power of Deep Blue, has no problem playing the sacrifice itself in about a minute. So I did learn something useful from Garry that day. The chess chip needed to recognize that certain piece configurations, in conjunction with open lines against the enemy king, could be very powerful in a king hunt.

 

 

第9章

CHAPTER 9

抚养婴儿

Bringing Up the Baby


选择芯片供应商

Picking the Chip Vendor

在卡内基梅隆大学,我从来没有考虑过为项目选择芯片供应商的问题。只有一种选择,MOSIS,MOS 实施服务。MOSIS 通过多个项目分摊成本,为学术界提供低成本芯片。NSF(国家科学基金会)和 DARPA 为大学 MOSIS 项目提供资金,通常学生可以免费制造芯片。

At Carnegie Mellon, I never thought about the problem of choosing a chip vendor for the project. There was only one choice, MOSIS, the MOS Implementation Service. MOSIS provided low cost chips to the academic world by sharing the cost among multiple projects. NSF (National Science Foundation) and DARPA provided the funding for university MOSIS projects, and usually students had their chips fabricated for free.

为了使用 MOSIS 服务,我必须自己进行芯片的定制布局。1985 年设计的包含 36,000 个晶体管的国际象棋走棋生成器的每个晶体管都是在计算机上手工绘制的。我还手工布线了芯片上的每根电线。1986 年,当我的日本朋友西泽 Teiji Nishizawa 在参加会议的途中访问卡内基梅隆大学时,他对我能够在 3 微米 CMOS 芯片上封装尽可能多的晶体管感到惊讶。当时的商业自动布局工具显然做得不太好。我只是不知道它有多穷。

To use the MOSIS service, I had to do the custom layout of the chip myself. Every single transistor for the 36,000-transistor chess move generator designed in 1985 was drawn by hand on a computer. I also hand routed every single wire on the chip. In 1986, when Teiji Nishizawa, my Japanese friend, visited Carnegie Mellon on his way to a conference, he was surprised that I managed to pack as many transistors as I did on a 3-micron CMOS chip. The commercial automatic layout tools at the time apparently did not do a very good job. I just had no idea how poor it was.

IBM 研究中心拥有一套优秀的内部 IC(集成电路)设计工具,据说比商业工具更好。IBM 微电子部门也拥有最先进的 IC 工艺,而且现在仍然如此。问题是,当我们刚加入 IBM 时,IBM 微电子公司只是为 IBM 内部批量生产而设立的。它并不是为了提供硅代工服务而设立的。今天,IBM 微电子公司提供了最好的(即使不是最好的)服务之一如果您想使用代工厂构建最快的微处理器,请致电 IBM 微电子公司。然而,在 1989 年,我遇到的所有 IBM 研究人员都建议我寻找外部芯片供应商。

IBM Research had an excellent set of internal IC (Integrated Circuit) design tools that were said to be better than the commercial tools. IBM Microelectronics Division also had the most advanced IC processes, and still does. The problem was that when we started at IBM, IBM Microelectronics was set up for volume IBM internal production only. It was not set up to provide a silicon foundry service. Today IBM Microelectronics provides one of the best, if not the best service—if you want to build the fastest microprocessors using a foundry, you call IBM Microelectronics. In 1989, however, all the IBM researchers I met suggested that I find an outside chip vendor instead.

在美国商业硅代工厂中,LSI Logic 和 VLSI Technology 是两个主要候选者。有了两家商业硅代工厂,我不再需要自己进行布局。我所需要做的就是设计电路,我们选择的供应商将使用他们自己的内部自动布局工具完成物理布局。我最初的想法是“这会很容易。” ChipTest和Deep Thought所使用的棋子生成器芯片的电路设计只花了一个月的时间。好吧,我还必须加上用于验证芯片的一个月时间。那还只是两个月。我推断:“新芯片应该不会超过一两年。”

Among the US-based commercial silicon foundries, LSI Logic and VLSI Technology were the two main candidates. With both commercial silicon foundries, I no longer had to do the layout myself. All that I needed to do was to design the circuit, and our chosen vendor would complete the physical layout using their own in-house automatic layout tools. My initial thought was “This is going to be easy.” The circuit design for the move generator chip used in ChipTest and Deep Thought took only a month. Okay, I would have to add in the one-month time used to verify the chip as well. That was still only two months. I reasoned, “The new chip should take no more than a year or two.”

经过一番调查,我倾向于超大规模集成电路技术。20 世纪 90 年代初,LSI Logic 将更多精力集中在“门阵列”业务上。在门阵列芯片中,所有晶体管都是预制的,并且通过在最后步骤中创建金属互连层来定制芯片。门阵列方法减少了从发送设计到接收芯片的时间,但代价是电路密度和设计灵活性。主要的替代方案是“标准单元”方法,它提供比门阵列方法更高的电路密度(通过仅在需要的地方制造晶体管)。然而,标准单元芯片的布局可以以与门阵列芯片大致相同的方式实现自动化。此外,对于需要大量片上存储器的芯片,标准单元芯片可提供更好的密度和更高的性能。VLSI Technology 拥有更成熟的标准单元技术,但最关键的论点是他们承诺根据我的博士论文中给出的设计,布置定制电路来检测重复的国际象棋位置 - 当然,要花一大笔钱。我将为 VLSI Technology 提供定制电路的详细晶体管级原理图。

After some investigation, I was leaning towards VLSI Technology. In the early 1990s LSI Logic was concentrating more on the “gate array” business. In a gate array chip, all the transistors are prefabricated and the chip is customized by creating the metal interconnection layers at the last steps. The gate array approach reduces the time from sending out the design to receiving the chip, but at the cost of circuit density and design flexibility. The main alternative is the “standard-cell” approach, which gives higher circuit density (by fabricating the transistors only where they are needed) than achievable with the gate array approach. The layout of the standard cell chips, however, could be automated in more or less the same way as the gate array chips. Furthermore, for chips that require substantial on-chip memory, standard cell chips offer much better density, and improved performance. VLSI Technology had a more mature standard cell technology, but the clinching argument was their promise to lay out the custom circuit for detecting repeated chess positions, based on the design given in my PhD dissertation—for a tidy sum of money, of course. I would provide VLSI Technology with detailed transistor level schematics for the custom circuit.

在评估 VLSI Technology 的工具期间,我对简化的国际象棋移动生成器的一格进行了试验布局。试验布局是针对 0.8 微米 CMOS 工艺的,它的密度应该是我在 1985 年完成的旧移动生成器中使用的 3 微米 CMOS 工艺的大约 14 倍。自动工具生成的试验布局仅约为密度是我在 3 微米工艺中手工布局的两倍。我封装晶体管的能力大约是自动工具的七倍。这是个好消息。坏消息是,我必须使用的自动工具在封装晶体管方面比我差大约七倍。

During the evaluation of the tools from VLSI Technology, I did a trial layout of one square of a simplified chess move generator. The trial layout was for a 0.8-micron CMOS process, which should have been about fourteen times denser than the 3-micron CMOS process that I used for the old move generator done in 1985. The trial layout generated by the automatic tool was only about twice as dense as my hand layout done in the 3-micron process. I was about seven times better at packing the transistors than the automatic tool. That was the good news. The bad news was that the automatic tool that I had to use was about seven times worse at packing the transistors than I was.

自动布局的包装密度较差是我决定对逻辑进行自定义布局以检测重复的国际象棋位置的主要原因。我认为这个决定是必要的,但这是一个糟糕的决定。今天我不会做出同样的决定。重复检测逻辑的定制设计中断了正常的设计流程,增加了设计的新复杂性,并显着延迟了芯片设计。我花了半年多的时间才使重复检测器的晶体管级逻辑正确。重复检测器也是我所涉及的交通事故的间接原因。

The poor packing density of automatic layout was the main reason for my decision to do a custom layout of the logic for detecting repeated chess positions. I thought the decision a necessary one, but it was a bad one. I would not make the same decision today. The custom design of the repetition detection logic interrupted the normal design flow, added new complexity to the design and delayed the chip design significantly. I spent over half a year just getting the transistor level logic correct for the repetition detector. The repetition detector was also the indirect cause of a traffic accident in which I was involved.

1993 年春末,也就是我们丹麦之行几个月后,我向 VLSI Technology 位于新泽西州的办公室提交了定制重复检测器的最终设计数据。在我回来的路上,一个粗心的司机把他的车开进了一条小路的停车标志,就在我前面大约二十米的地方。这是一次改变生活的经历,但我不建议任何人尝试。当另一辆车突然冲到我前面的路上时,我以为我要死了,项目1就结束了。我设法充分转向我租来的车,避免直接撞上另一辆车。如果他看到我并及时停下来,事故就可以避免了。结果他不顾路况继续前行,结果发生了一次擦撞。我还活着,但我的脸被撞伤了,眼镜被安全气囊撞碎了。我的脸部和左臂也有撕裂伤。另一辆车内的乘客没有受伤。在附近的一家医院,我发现我的左眼角膜有一个严重的割伤,我不得不戴一个星期的眼罩。由于眼睛的问题,我大约一个月没有真正恢复认真的工作2。当我回到工作岗位时,由于长时间的休息以及可能是事故造成的创伤,我很难理解评估硬件的现有设计。花费大约一小时后一周回顾了早期的设计,我放弃了并决定从头开始重做评估硬件。我已经在评估硬件上花了大约半年的时间,这一切都被浪费了。

In the late spring of 1993, a few months after our Denmark trip, I delivered the final design data for the custom repetition detector to VLSI Technology’s office in New Jersey. On my way back, a careless driver ran his car through a stop sign from a side road, only about twenty meters in front of me. It was a life-altering experience, but I would not recommend anyone to try it. When the other car suddenly burst onto the road in front of me, I thought that I was going to die and it would be the end of the project1. I managed to swerve my rental car sufficiently to avoid hitting the other car straight on. Had he seen me, and stopped in time, an accident would have been avoided. As it was, he went on without any regard to the road condition and a glancing collision took place. I was alive, but with the face bruised and cut by my glasses that were shattered in the impact with the air bag. I also had lacerations on my face and left arm. The occupants in the other car were uninjured. At a nearby hospital it was found that the cornea of my left eye had a nasty cut, and I had to wear an eye patch for a week. I did not really get back to serious work for about a month because of the problem with my eye2. By the time I got back to work, as a result of the prolonged break and possibly the trauma from the accident, I had a hard time understanding the existing design of the evaluation hardware. After spending about a week going over the earlier design, I gave up and decided to redo the evaluation hardware from scratch. I had already spent about half a year on the evaluation hardware which was all wasted.

情况变得更糟了。1994 年末,当有关深蓝和加里·卡斯帕罗夫之间潜在比赛的讨论开始时,深蓝国际象棋芯片的第一个版本仍在设计中。一年多后,VLSI Technology 的定制重复检测器仍然没有出现。随着时间的推移,我们决定从设计中删除定制的重复检测器。因此,设计探测器所花费的六个月也付诸东流。总的来说,创建自定义重复检测器的一个错误决定使项目延迟了一年。这是一个代价高昂的决定,不仅浪费了时间,而且还大大削弱了国际象棋机器的性能。虽然检测重复乍一看似乎不是很重要,但正如您稍后会看到的那样,它可以对国际象棋程序的走位产生非常微妙的影响。

It got worse. In late 1994, when discussions about a potential match between Deep Blue and Garry Kasparov were initiated, the first version of the Deep Blue chess chip was still being designed. There was still no sign of the custom repetition detector from VLSI Technology after over a year. With time getting short, the decision was made to drop the custom repetition detector from the design. So the six months spent on designing the detector went down the drain as well. In total, the one bad decision to create a custom repetition detector delayed the project by one year. A pricey decision, not just in terms of the wasted time, but also in terms of a vastly weakened chess machine. While detecting repetition does not seem very important at first sight, as you will see later, it can have a very subtle influence on a chess program’s positional play.

大约在这个时候,VLSI Technology 决定将工作转移到他们的波士顿办公室。几个月后,当我向 VLSI 波士顿办事处交付完整网表的第一个版本(没有重复检测逻辑)时,我看到了自定义重复检测器的部分布局。布局大约比我预期的大五倍,超过 40% 的面积被两个巨大的线束从一侧交叉到另一侧占据。自动工具会发现这样一个事实:两半可以像梳子一样交错,并且交叉线应该占据很小的面积。因此,自定义布局在某种程度上变得比自动布局更糟糕。VLSI 新泽西办事处的工程师被告知电线应该交错,但不知何故,该信息从未传达给布局技术人员。即使我们有时间等待其完成,我们也永远不会接受原来的自定义布局。

At about this time VLSI Technology decided to move the work to their Boston office. A few months later, when I delivered the first version of the complete netlist to VLSI’s Boston office, sans the repetition detection logic, I saw the partial layout of the custom repetition detector. The layout was about five times bigger than I expected, with over 40 percent of the area occupied by two giant wire bundles crossing each other from one side to the other. An automatic tool would have caught the fact that the two halves could be interleaved like a comb, and the crossing wires should occupy very little area. So the custom layout somehow became much worse than an automatic layout. The engineer at VLSI’s New Jersey office had been told that the wires should be interleaved, but somehow the information was never relayed to the layout technician. We never would have accepted the custom layout as it was, even if we had the time to wait for its completion.

今天,如果我设计另一个国际象棋芯片或任何其他芯片,我会以完全不同的方式处理芯片供应商的决策。自 20 世纪 80 年代末以来,半导体行业发生了很大变化。然后,芯片设计人员通常使用芯片供应商特定的工具,通过原理图设计电路。从 20 世纪 90 年代初开始,芯片设计人员转向基于语言的逻辑综合工具,用硬件描述语言描述芯片,并依靠自动逻辑综合工具来创建网表。该芯片的设计可以不考虑特定的芯片供应商。设计人员可以在设计过程的最后阶段选择芯片供应商,因为更换芯片供应商只会造成很小的损失。唉,当国际象棋项目开始时,我并没有这种奢侈。当综合工具被广泛使用时,IBM 陷入了严重的财务困境,没有钱购买新工具。尽管在项目接近尾声时有更好的设计工具可用,但深蓝国际象棋芯片几乎完全使用电路原理图进行设计。由于使用特定于供应商的工具,我们在芯片技术方面也没有太多选择。1997 年加里·卡斯帕罗夫 (Garry Kasparov) 与深蓝 (Deep Blue) 之间的复赛是使用 0.6 微米 CMOS 芯片进行的,尽管业界已经转向 0.35 微米 CMOS 技术并超越3 微米。无论如何,选择 VLSI 技术而不是 LSI 逻辑的理由如今已不再有效。特别是,我还会避免像瘟疫这样的任何自定义布局,除非我可以完全控制它。

Today, if I design another chess chip or any other chip, I would approach the chip vendor decision quite differently. The semiconductor industry has changed quite a bit since the late 1980s. Then, chip designers designed the circuit with schematics, often using chip vendor specific tools. From the early 1990s, chip designers turned to language-based logic synthesis tools, describing the chip in hardware description languages and relying on automatic logic synthesis tools to create the netlist. The chip could be designed without a specific chip vendor in mind. The designer could pick the chip vendor very late in the design process since there is only a small penalty to switch the chip vendor. Alas, when the chess project was started, I did not have this luxury. By the time synthesis tools became widely used, IBM was in serious financial trouble and there was no money to acquire the new tools. The Deep Blue chess chips were designed almost entirely using circuit schematics even though there were far better design tools available near the end of the project. We also did not have many choices on the chip technology front as a result of using vendor-specific tools. The 1997 rematch between Garry Kasparov and Deep Blue was played using 0.6-micron CMOS chips, even though the industry had already moved on to 0.35 micron CMOS technology and beyond3. Anyway, the reasons for choosing VLSI Technology over LSI Logic would no longer be valid today. In particular, I would also avoid any custom layout like the plague, unless I have full control over it.

当我们做出决定时,选择 VLSI 技术可能与任何其他选择一样好。在最终计票中,VLSI 交付了新协议的结尾。所以一切都顺利了。后来在设计工具上有更好的选择,但考虑到当时 IBM 的预算紧张,我们无法做出选择。当 IBM 的财务回到正轨时,设计已接近完成,而更改工具为时已晚。

At the time that we made the decision, choosing VLSI Technology was probably as good as any other choice. In the final count, VLSI delivered their end of the new agreement. So everything did turn out all right. There were better choices on the design tools later on, but we were in no position to make the choice, given the tight IBM budget at the time. By the time IBM’s finances were back on track, the design was already close to completion and it was too late to change the tools.

墨菲定律

Murphy’s Law

1994 年末,IBM、加里·卡斯帕罗夫 (Garry Kasparov) 和 ACM(计算机协会)(这场大型比赛的主要批准组织)之间进行了认真的谈判。比赛定于 1996 年 2 月举行,与庆祝活动同时进行。纪念 ACM 成立 50 周年。深蓝国际象棋芯片的设计已接近完成,但验证其正确性的任务依然艰巨。默里和乔同意放弃他们正在做的事情并帮助芯片验证。Deep Thought II 被关闭并断开连接,因为我们需要它的主机,这是我们唯一具有足够主内存来运行芯片模拟的机器。我们原本希望在一些测试游戏中重新开启 Deep Thought II新的国际象棋机器何时投入运行;否则,对我们来说,《深思二号》就已经不复存在了。

Serious negotiations between IBM, Garry Kasparov, and the ACM (Association for Computing Machinery), which was the main sanctioning organization for the big match, were taking place in late 1994. The match was to be held in February 1996 to coincide with the celebration marking the 50th anniversary of ACM. The design of the Deep Blue chess chip was near completion, but there was still the huge task of verifying its correctness. Murray and Joe agreed to drop what they were doing and help with the chip verification. Deep Thought II was turned off and disconnected because we needed its host computer, which was our only machine with sufficient main memory to run the chip simulation. We were expecting to turn Deep Thought II back on for a few test games when the new chess machine became operational; otherwise, as far as we were concerned, Deep Thought II was no more.

ICCA(国际计算机象棋协会)每三年举办一次世界计算机象棋锦标赛。他们联系IBM赞助这项将于1995年5月在香港举行的赛事。我们在1989年赢得了该赛事,但没有参加1992年。如果完全由我决定,我们可能会再次跳过该赛事。默里、乔和我已经忙于芯片工作,因为确保我们为加里·卡斯帕罗夫做好准备似乎更为重要。

The ICCA (International Computer Chess Association) held a World Computer Chess Championship tournament triennially. They approached IBM to sponsor the event, to be held in Hong Kong in May 1995. We had won the event in 1989, but did not participate in 1992. If it had been entirely up to me, we probably would have skipped the event again. Murray, Joe and I were already swamped with the chip work as it seemed far more important to make sure that we would be ready for Garry Kasparov.

然而,我们的经理 CJ Tan 有不同的想法。他认为这次锦标赛是为与加里的比赛增加宣传的机会,因此安排获得锦标赛的资金并同意参加。新的国际象棋芯片在香港比赛中使用的可能性很小,但很快就发现,到那时它甚至还没有准备好制造。唯一的选择就是与 Deep Thought II 竞争。

However, our manager, C J Tan, had a different idea. Seeing the championship as a chance to add publicity to the match with Garry, he arranged to get the funding for the championship and agreed to participate. There was a slim chance that the new chess chip might be available for the Hong Kong event, but it soon became clear that it would not even be ready for fabrication by then. The only other choice was to compete with Deep Thought II.

Jerry Present,我们当时的通讯人员,想使用 Deep Blue Prototype 这个名称,而不是 Deep Thought II。我想使用“Deep Thought II”这个名字。我的推理如下。新机器“深蓝”的有效搜索速度​​将比“深思二号”至少快100倍,而且“深蓝”对国际象棋中的位置概念的掌握也会好得多。将深蓝与深思 II 进行比较就像将太阳与月亮进行比较一样。好吧,我有点夸张,但考虑到在深蓝国际象棋芯片上进行的更为复杂的国际象棋评估计算,计算能力的差异大约为千比一的数量级。我对所做的事情感到自豪,并且不希望任何东西与深蓝相关,除非它确实使用了新芯片。Deep Thought II 仍然使用我在 1985 年设计的国际象棋筹码。在我看来,Deep Thought II 是一只即将灭绝的恐龙。它不会篡夺我们新的百万吨级太阳能爆炸器的名称,即使只是部分篡改。

Jerry Present, our communications person at the time, wanted to use the name Deep Blue Prototype instead of Deep Thought II. I wanted to use the name Deep Thought II. My reasoning was as follows. Deep Blue, the new machine, would be at least 100 times faster than Deep Thought II in effective search speed, and furthermore, Deep Blue would have a far better grasp of positional concepts in chess. Comparing Deep Blue with Deep Thought II would be like comparing the sun with the moon. Well, I exaggerate a little bit, but the difference in computation power was roughly of the order of a thousand to one, taking into account the far more complicated chess evaluation computation carried out on the Deep Blue chess chips. I was proud of what had been done, and did not want anything to be linked to Deep Blue unless it did use the new chips. Deep Thought II was still using the chess chips that I designed back in 1985. As far as I was concerned, Deep Thought II was a dinosaur about to become extinct. It was not going to usurp the name of our new megaton solar blaster, even if only partially.

然而,杰瑞说服 CJ 坚持使用“深蓝原型”这个名字。通信部门希望能够在与加里·卡斯帕罗夫(Garry Kasparov)比赛时说“深蓝”是卫冕世界计算机象棋冠军“深蓝原型”的继承者。在“不太可能”的情况下,Deep Thought II 没有赢得冠军,那么它就不是真正的冠军。无论如何,深蓝都在玩。问题是,考虑到瑞士锦标赛的赛制,“不太可能”的可能性很大。尽管事实上 Deep Thought II 在实验室测试中以大约三比一的优势击败了我们最接近的竞争对手(或者说计算机与计算机游戏的评分差异为 200 分),但我们赢得冠军的机会只有大约 50-50。参赛队伍有二十四支,只进行五轮比赛。在如此少的比赛中,运气对于赢得比赛起到了非常重要的作用。我们是最强的,不过比他们低一级的大概有五六支队伍,还有三五支队伍只比他们稍弱一些。默里、乔和我这支球队都知道在这种赛制下我们真正的机会。我们的第一任主教练兰迪·莫利克(Randy Moulic)也知道这种可能性,因为他经历过 1990 年的 ACM 锦标赛,但我们没有彻底获胜。但CJ只和我们一起参加了1994年的ACM锦标赛。在那场比赛中,Deep Thought II 彻底赢得了冠军,尽管纽约的一场突如其来的暴风雨导致实验室断电,我们输掉了一场比赛4

Jerry, however, persuaded C J to stick with the name Deep Blue Prototype. The Communications Department would like to be able to say that Deep Blue was the successor to the reigning World Computer Chess Champion, Deep Blue Prototype, at the time of the match with Garry Kasparov. In the “unlikely” case that Deep Thought II did not win the championship, then it was not really Deep Blue playing anyway. The problem was that, given the Swiss tournament format of the championship, the “unlikely” was highly probable. Despite the fact that Deep Thought II beat our closest competitors by about a three to one margin (or 200 rating points difference in computer vs. computer play) in lab tests, our chance of winning the championship was only about 50–50. There were twenty-four competing teams, and only five rounds of games would be played. With so few rounds, luck played a very important part in winning the tournament. We were the strongest, but there were about five or six teams just one class below, and another three to five teams only slightly weaker than they. The team, Murray, Joe, and I, all knew about our real chances given the format. Randy Moulic, our first manager, also would have known about the odds, having gone through a 1990 ACM Championship that we did not win outright. But C J had only been through the 1994 ACM Championship with us. In that one event, Deep Thought II won the title outright, even though we forfeited one game when a sudden storm in New York knocked out the electric power to the lab4.

Deep Thought II 一直处于关闭状态,直到香港活动前一个月左右。穆雷花了很少的时间为冠军做准备。新芯片的最后验证工作还在进行中,我们没有太多的空闲人力。我们必须腾出穆雷的一些时间,因为他是国际象棋专家。除了 ChipTest 的第一次亮相之外,我们在准备计算机国际象棋比赛时从未如此漫不经心。然后,我们不在乎我们是否获胜,而对于《Deep Thought II》的最后一次演出,我们主要关心的是与加里的约会,还有九个月的时间。

Deep Thought II remained turned off until about a month before the Hong Kong event. Murray spent a small amount of time preparing it for the championship. The final verification work on the new chip was still going on, and we did not have much manpower to spare. We had to free up some of Murray’s time, as he was the chess expert. We had never been so cavalier in our preparation for a computer chess event, save for ChipTest’s first outing. Then, we did not care whether we won or not, and for this last outing of Deep Thought II, our main concern was keeping the date with Garry, nine months away.

似乎当你准备最不充分时,事情通常就会出错。正如墨菲定律所预测的那样,一切可能出错的事情很快就会出错。

It seems that things usually go wrong when you are least prepared. As Murphy’s Law predicts, everything that could go wrong soon went wrong.

我和默里与 CJ 一起飞往香港担任现场操作员。乔留下来监视“深思二号”。这是我自1985年开始国际象棋项目以来第一次回到亚洲,也是我第一次去香港。

Murray and I, traveling with C J, flew to Hong Kong to be the on-site operators. Joe stayed behind to watch over Deep Thought II. It was the first time since the start of the chess project in 1985 that I had been back to Asia, it was also my first trip to Hong Kong.

当穆雷和我前往比赛场地香港中文大学时,潜在麻烦的第一个迹象就显现出来了。无法连接到美国的互联网相当有效;就此而言,与欧洲的联系也没有。香港的三所主要大学使用一个公共运营商来进行所有互联网连接,但其容量有限。活动开始前,拥有远程计算机的团队已尝试从自己的国家连接到香港的互联网,效果良好。但使用互联网连接离开香港几乎是不可能的,至少在比赛期间是这样。看起来远程团队可能必须支付非常昂贵的长途电话费。

The first hint of potential trouble exhibited itself when Murray and I went to the playing site, the Chinese University of Hong Kong. The Internet connection to the United States was not quite working; for that matter, neither was the connection to Europe. The three main universities in Hong Kong used a common carrier, which had a limited capacity, for all their Internet connections. Before the event, the teams with computers on remote sites had tried out Internet connections into Hong Kong from their home countries, and they appeared fine. But getting out of Hong Kong using an Internet connection was almost impossible, at least during the tournament. It looked as though the remote teams might have to pay very expensive long distance phone bills.

碰巧的是,在锦标赛的前一天,IBM 香港公司宣布为香港提供 IBM 互联网服务。尽管连接似乎时不时地掉线,但这项服务还是可用的。这可能是本地电话服务的问题,因为我们拨打长途时也遇到了问题。(1997年,我回到香港,通过IBM互联网连接进行深蓝演示,进展顺利。所以我们在1995年遇到的沟通问题可能根本不是互联网连接的问题。)电话线问题变成了这是 Deep Thought II 上一场比赛的一个重要因素。

As it happened, the day before the championship IBM Hong Kong had announced an IBM Internet service for Hong Kong. This service was usable although the connection seemed to drop from time to time. It might have been a problem with the local phone service, since we were having problems when we dialed long distance as well. (In 1997, I was back in Hong Kong doing a Deep Blue demo via the IBM Internet connection, and it went smoothly. So the communication problem we had in 1995 might not be with the Internet connection at all.) The phone line problem became an important factor in the last game played by Deep Thought II.

我们的第一轮对手是 Star Socrates,这是麻省理工学院的一个并行国际象棋程序,运行在一台价值数百万美元的超级计算机上——这台机器比 Deep Thought II 运行的工作站贵大约一百倍。明星苏格拉底被认为是我们最危险的对手。它的搜索速度至少可以与深思二号相媲美。当 Deep Thought II 前一年玩《苏格拉底之星》时,我第一次看到一个程序实际上“搜索超过”了 Deep Thought II——它报告的搜索深度比 Deep Thought II 的还要大。然而《深思二号》却轻松地超越了它。一年会有什么不同吗?Deep Thought II 已经下线半年多了,麻省理工学院的人可能也没有闲着。1994年,Deep Thought II曾采用常规的国际象棋开局,但由于香港赛的开局准备时间只有一个月,穆雷选择采用不规则的开局,将这两个程序从开局记录中剔除。这场比赛变成了一场原始棋力的较量。我对穆雷的开局选择不太满意;我们有白棋,但我们在开局时的处境稍差。然而,《深思二号》很快就战胜了明星苏格拉底。

Our first round opponent was Star Socrates, a parallel chess program from MIT, running on a multi-million-dollar supercomputer—a machine about a hundred times more expensive than the workstation that Deep Thought II ran on. Star Socrates was considered our most dangerous opponent. Its search speed was at least comparable to Deep Thought II. When Deep Thought II played Star Socrates the year before, I saw for the first time a program that actually “out-searched” Deep Thought II—it was reporting search depths larger than Deep Thought II’s. Yet Deep Thought II had outclassed it easily. Would one year make a difference? Deep Thought II had been off line for over half a year, and the MIT folk were probably not sitting idle. In 1994, Deep Thought II had played a regular chess opening, but with only one month of opening preparation for the Hong Kong event, Murray opted to play an irregular opening to take both programs out of their opening books. The game became a contest of raw chess playing strength. I was not too happy with Murray’s opening choice; we had White, and yet we had a slightly worse position out of the opening. Deep Thought II, however, soon overpowered Star Socrates.

接下来的两轮,深思II继续打出不规则的开局;我们赢了两场比赛,并以一场完整的比分领先。我继续向穆雷抱怨糟糕的开局。回想起来,我应该保持沉默。穆雷选择打不规则的开局是出于必要——没有时间做适当的开局准备。Deep Thought II 是一名比场上其他选手更好的选手,通过从他们的开局书中剔除其他程序,穆雷给了我们的对手一根更长的绳索来吊死自己。

For the next two rounds, Deep Thought II continued to play irregular openings; we won both games and led the field by one full game. I continued to mumble to Murray about the bad openings. In retrospect, I should have kept quiet. Murray chose to play irregular openings out of necessity—there was no time to do proper opening preparation. Deep Thought II was a better player than the rest of the field, and by taking other programs out of their opening books, Murray gave our opponents a longer rope with which to hang themselves.

我们的第四个对手是WCess,是我们更强硬的对手之一。《深思II》这次打出了正常的开局。当这两个项目刚出版时,情况仍然如此。Deep Thought II 很快就找到了赢得棋子的方法并占据了制胜位置,但 WCess 却有一些反击。Deep Thought II 下了一个糟糕的棋,其地位恶化了。几乎就在这之后,Deep Thought II 发现 WCess 有一种绘制游戏的方法。WChess 此时并未看到平局,但它下了正确的棋步。Deep Thought II突然发现自己的主教已经快被困住了,不得不放弃材料来保住主教。WCess 即将在材料方面取得进展。事情的发展很快。此时,WCess 看到了平局,并且很快也看到了它会领先。对我们来说幸运的是,Deep Thought II 在其糟糕的举动之前有足够的吸引力,尽管有糟糕的举动,我们还是能够保持平局。这场比赛之后,我对我们决定从新的国际象棋芯片的设计中放弃重复检测器感到有点不安。如果 Deep Thought II 有能力检测硬件中的重复,它就会发现糟糕的棋步会导致什么结果,并下出更好的棋步。

Our fourth opponent was WChess, one of our tougher opponents. Deep Thought II played a normal opening this time. The position was even when both programs came out of their opening books. Deep Thought II soon found a way to win a pawn and had a winning position, but WChess had some counter play. Deep Thought II made one bad move and its position deteriorated. Almost immediately after this, Deep Thought II found that WChess had a way to draw the game. WChess did not see the draw at this point, but it played the right moves. Deep Thought II suddenly found its bishop was close to being trapped and it had to give up material to keep the bishop. WChess was about to go ahead in material. It was a fast turn of events. At this point WChess saw the draw, and soon also saw that it would be ahead. Luckily for us, Deep Thought II had enough pull before its bad move, and despite the bad move we were able to hold the draw. After this game, I was a little bit uneasy about our decision to drop the repetition detector from the design of the new chess chip. If Deep Thought II had had the ability to detect repetition in hardware, it would have found where the bad move would have led and played a better move instead.

尽管如此,我们仍然领先半场,而我们上一轮的对手是弗里茨,对我们来说是一个轻松的对手。弗里茨的强项是它的战术,但深思二号是一个更好的战术家;事实上,在 IBM 的赛前测试中,它对阵 Fritz 的十场比赛中大约有九场获胜。回酒店的路上我已经筋疲力尽了。前几天,我一直在抱怨开局不好。到了这个时候,尤其是在第四轮的结果不太好之后,我认为不规则的开局方式是一个非常好的主意。考虑到弗里茨的位置感相对较弱,不规则的开局应该会很有效。这一次,当我应该说话时,我保持沉默。那天晚上我入睡时,我以为穆雷会为弗里茨准备一些不规则的开场白。第二天带来了很大的惊喜。

Nonetheless, we still led the field by half a game, and our last round opponent was Fritz, an easy opponent for us. Fritz’s strong point was its tactics, but Deep Thought II was a far better tactician; in fact, it won about nine out of ten games against Fritz in our pre-tournament tests back at IBM. I was exhausted on the way back to our hotel. The previous few days, I had been complaining about the bad openings. By this time, especially after the not so good result in round 4, I was of the opinion that the irregular opening approach had been a very good idea. Given the relatively weak positional sense of Fritz, an irregular opening should work really well. For once, I kept quiet when I should have spoken up. When I went to sleep that night, I was expecting Murray to cook up some irregular opening for Fritz. The next day brought a big surprise.

穆雷早早来到比赛现场并与一些人聊天其他参与者的。有人告诉他,弗里茨的开局书很糟糕,并且在开局书里就输掉了一场比赛。穆雷认真对待了这个评论,并决定对弗里茨使用正常的开局。毕竟,如果弗里茨在我们演奏不规则的开局时是小菜一碟,那么如果我们演奏真正的开局,那就更容易了,特别是考虑到弗里茨的开局书很糟糕。穆雷赛前没有告诉我他的决定。《深度思考 II》以国王兵开局开局,我疑惑地看了穆雷一眼。弗里茨在西西里防御中以尖锐的战线予以回应。弗里茨开篇书的作者巧妙的“换位”(故意将开局动作与正常的演奏顺序交换)将《深度思考 II》从其自己的开篇书中取出。Deep Thought II 有四种主要的移动选择,其中两种是易位移动,通常将国王从中心转移到更安全的侧翼位置。这两种易位动作(可能是弗里茨开篇书的作者所设计的)在这个位置上都很糟糕。Deep Thought II 认为这个位置对自己有利,因此想要占领城堡。穆雷的脸色白了。对我们来说,这个位置看起来很不稳定。这是《深思二》开卷中的第一个动作,并且在动作上花费了额外的时间。也许有人希望“深思二号”会走更安全的棋。突然我们就断了联系。据在霍桑实验室观看比赛的乔说,《深度思考 II》确实转向了另一种走法。但在掉线之前,这个新动作从未出现在我们香港的屏幕上,直到比赛结束后我们才知道。

Murray went to the playing site early and chatted with some of the other participants. One told him that Fritz had a horrible opening book, and had lost one of its games right out of the opening book. Murray took the comment seriously and decided to use a normal opening against Fritz. After all, if Fritz had been a piece of cake when we played irregular openings, it would be even easier if we played a real opening, especially given that Fritz had a bad opening book. Murray did not tell me about his decision before the game. Deep Thought II opened with a king pawn opening, and I gave Murray a quizzical glance. Fritz replied with a sharp line in the Sicilian Defense. A clever “transposition” (deliberate swapping of opening moves from their normal playing sequence) on the part of the author of the Fritz opening book took Deep Thought II, out of its own opening book. There were four main move choices for Deep Thought II, and two of them were castling moves that normally take the king out of the center into the safer wing positions. Both castling moves, probably by design of the author of the Fritz opening book, were bad in this position. Deep Thought II assessed the position as favorable for itself and wanted to castle. Murray’s face turned white. The position looked precarious for us. It was the first move out of the opening book for Deep Thought II, and it was spending extra time on the move. Perhaps there was some hope that Deep Thought II would play a safer move. Suddenly we got disconnected. According to Joe, who was watching the game from our lab in Hawthorne, Deep Thought II did switch to an alternative move. But the new move never showed up on our screen in Hong Kong before the line drop, and we did not know about it until after the game.

麻省理工学院的明星苏格拉底团队正在使用一个程序,该程序允许他们重新连接到他们的国际象棋程序,而无需重新启动。如果我们使用相同的程序,我们就会重新连接,找到 Deep Thought II 已经在其内部棋盘上下的新棋,并在真正的棋盘上下棋。事实上,我们不得不重新启动 Deep Thought II,现在花在这个位置上的时间更少了。更糟糕的是,它是在没有先前计算帮助的情况下“冷”启动的,因此有效速度可能慢了两倍,“深度思考 II”再也找不到更合理的替代举措。

The MIT Star Socrates team was using a program that allowed them to reconnect to their chess program without the need to restart. Had we used the same program, we would have reconnected, found the new move already played by Deep Thought II on its internal board, and played it on the real chessboard. As it was, we had to restart Deep Thought II, which now had less time to spend on the position. Compounded by the fact that it was started “cold” without the help of previous calculations and thus probably a factor of two slower effectively, Deep Thought II could no longer find the more reasonable alternative move.

在弗里茨采取下一步行动后,《深思二号》的评价跌至重复平局。所以这也是一个重复检测器问题!如果 Deep Thought II 有一个硬件重复检测器,它就可以避免糟糕的易位移动,或者没有掉线。对于从新芯片中删除硬件重复检测器的决定,我的感觉更加糟糕。

After Fritz’s next move, the evaluation of Deep Thought II dropped to a repetition draw. So it was a repetition detector problem as well! If Deep Thought II had had a hardware repetition detector, it would have avoided the bad castling move, with or without the line drop. I was feeling even worse about the decision to drop the hardware repetition detector from the new chip.

我们对弗里茨的处境越来越糟糕,但并非毫无希望。我们的通讯线路仍然存在问题。我们得过且过,浪费了宝贵的时间,但我们还没有被抓。然后它击中了。Deep Thought II发现它想要下的棋步正在失败,并进入“恐慌时间”状态,它会花费额外的时间来寻找替代棋步。《纽约时报》国际象棋专栏作家、国际象棋大师罗伯特·伯恩 (Robert Byrne)也是 ICCA 的特邀嘉宾,他顺便过来低声说,“深度思考 II”必须下某一步棋才能保住局面。深思二号打算的举动并不是罗伯特所建议的举动。我们无助地看着 Deep Thought II 旋转着,试图找到一个好的替代方案。搜索似乎呈爆炸式增长——“深思二号”突然搜索了一棵大得多的树。一些替代举措被拒绝。罗伯特的举动能被及时发现吗?时钟滴答作响。时间到了。由于没有更好的办法,“深思 II”发挥了最初的意图。比赛实际上已经结束了。弗里茨(Fritz)没有发挥出最好的进攻动作,但足以赢得被淘汰的位置,因此我们放弃了“深思二号”(Deep Thought II)。弗里茨在击败明星苏格拉底后最终赢得了比赛,但明星苏格拉底在季后赛中自毁了。

Our position against Fritz was getting bad, but it was not hopeless. We continued to have problems with the communication line. We muddled along, losing precious time on our clock, but we were not busted yet. Then it hit. Deep Thought II found the move that it wanted to play, was losing, and entered into the “panic time” state where it would spend extra time to find an alternative move. Grandmaster Robert Byrne, the New York Times chess columnist who was the honored guest of ICCA, dropped by and whispered that Deep Thought II had to play a certain move to keep the position alive. The move intended by Deep Thought II was not the move that Robert suggested. We watched helplessly as Deep Thought II spun its wheels, trying to find a good alternative. The search appeared to be exploding—Deep Thought II was suddenly searching a much larger tree. Several of the alternative moves were rejected. Would Robert’s move be found in time? The clock was ticking. Time was up. Deep Thought II played its originally intended move as nothing better had come up. The game was effectively over. Fritz did not play the best attacking moves, but it was good enough to win the busted position, and we resigned for Deep Thought II. Fritz ended up winning the tournament after beating Star Socrates, when it self-destructed in their playoff game.

这次比赛教会了我们很多东西。从我的角度来看,最重要的是,如果我有另一次机会制作更新版本的国际象棋芯片,它必须有一个硬件重复检测器。还有其他教训。我们必须更加认真地对待开场准备工作,每当我们进行远程比赛时,我们都必须有良好的后备计划。我们应该能够像麻省理工学院团队一样重新连接到国际象棋程序,并且我们应该事先彻底测试通信线路。在与加里的比赛之前,我们无法解决硬件重复检测器问题,但我们当然可以解决通信问题。

The tournament taught us several things. From my point of view, the most important was that if I ever got another chance to make a newer version of the chess chip, it must have a hardware repetition detector. There were other lessons too. We had to take opening preparations far more seriously, and we must have good backup plans whenever we played remotely. We should be able to reconnect to the chess program in the same way as the MIT team, and we should test out the communication lines thoroughly beforehand. We could not do anything for the hardware repetition detector problem before the match with Garry, but we certainly could do something about the communication issues.

使用“深蓝原型”这个名字的决定也困扰着我们。几家电脑象棋供应商在描述他们在香港比赛中的结果时,将“深蓝原型”更名为“深蓝”。有人甚至说他们的程序击败了加里·卡斯帕罗夫(在闪电战游戏中)和深蓝(它从未玩过深蓝)。虽然,公平地说,他们即使我们使用 Deep Thought II 这个名字,供应商也可能会这样做。

The decision to use the name “Deep Blue Prototype” also came back to haunt us. Several computer chess vendors renamed “Deep Blue Prototype” to “Deep Blue” when describing their result from the Hong Kong event. One went as far as saying that their program beat both Garry Kasparov (in blitz games) and Deep Blue (it never played Deep Blue). Although, to be fair to them, the vendors probably would have done the same even if we had used the name Deep Thought II.

剩余四个小时

Four Hours to Spare

虽然我们放弃了在香港赛上使用新棋子的计划,但还有另一场比赛必须使用它。IBM 研究院计划在中国北京开设一个新的研究实验室,作为开业庆典的一部分,我们受邀与国际象棋女子世界冠军谢军进行了一场友谊赛。北京比赛定于1995年9月下旬进行。与加里·卡斯帕罗夫的比赛合同中有一项条款规定IBM可以在1995年10月31日或之前终止比赛合同。球队并没有明确告知我们必须在北京比赛中取得好成绩但是,考虑到时间的接近,这种联系是显而易见的。

Although we dropped the plan to use the new chess chip for the Hong Kong tournament, there was another event where it would have to be used. IBM Research was planning to open a new research lab in Beijing, China, and we were invited to play a friendly match with Xie Jun, the Women’s World Chess Champion, as a part of the opening celebration. The Beijing match was scheduled for late September 1995. A clause in the match contract with Garry Kasparov stated that IBM could terminate the match contract on or before 31 October 1995. The team was not told explicitly that we had to do well in the Beijing match but, given the time proximity, the link was clear.

1995 年 8 月下旬,VLSI Technology 给我们带来了一些令人不安的消息。第一批芯片均未通过制造测试。他们在九月初向我们发送了一些表现更好的芯片。与谢军的比赛原定于9月21日举行,距离比赛还有不到三周的时间。即使筹码已经通过,日程也非常紧张。

In late August 1995, VLSI Technology gave us some disturbing news. None of the chips in the first batch passed the manufacturing test. They sent us some of the better behaving chips in early September. The Xie Jun match was scheduled for September 21, less than three weeks away. It was a very tight schedule even if the chips had passed.

当我们从 VLSI Technology 得到消息时,Joe 已经开始将 Deep Thought II 软件移植到新硬件上。几周前我收到了新芯片的印刷电路板,并对它们进行了基本测试。现在,乔和我一起度过了一段愉快的时光。

At the time we got the news from VLSI Technology, Joe had started porting the Deep Thought II software to the new hardware. I had received the printed circuit boards for the new chips a few weeks earlier and had done the basic tests on them. Now, Joe and I were in for some fun time together.

对印刷电路板的基本测试只是检查电路板是否可以正确初始化。我准备了更高级的测试来测试主板上的芯片。没有一个芯片通过了最简单的高级测试,这些测试检查了芯片之间总线的连接性。问题可能出在主板上,也可能出在芯片上。板卡的基本测试仅使用 IBM RS/6000 工作站中 32 位 Microchannel 总线的 8 位。高级测试使用了全部 32 位来检查芯片操作。可能是24个未经测试的总线位有问题。使用欧姆表5,我检查了芯片之间的连接以及板上的微通道总线,发现电线在物理上是连接的,但不知何故它们连接在一起错误地。事实证明,RS/6000 工作站对 32 位的解释与印刷电路板的解释不同。用技术术语来说,这就是所谓的“字节交换”问题。32位可以分为4个字节,每个字节有8位。工作站假设 4 个字节以某种方式排序,而主板则假设 4 个字节以不同的方式排序。好吧,这并不令人愉快,但是可以通过在软件中重新解释 32 位来绕过这个问题。

The basic tests for the printed circuit boards simply checked whether the boards could be initialized properly. I had prepared more advanced tests for testing the chip with the board. None of the chips passed the simplest of the advanced tests that checked the connectivity of the bus wires to and from the chips. The problem could be with the board or it could be with the chips. The basic tests for the boards used only 8 bits of the 32-bit Microchannel bus in the IBM RS/6000 workstation. The advanced test used all 32 bits to check out the chip operations. It was possible that there were something wrong with the 24 untested bus bits. Using an ohmmeter5, I checked the connections between the chips, and the Microchannel bus on the board, and found that physically the wires were connected, but somehow they were connected incorrectly. It turned out that the RS/6000 workstation interpreted the 32 bits differently from the printed circuit board. In technical terms, this is the so-called “byte swapping” problem. The 32 bits can be divided into 4 bytes, with each byte having 8 bits. The workstation assumed that the 4 bytes were ordered in a certain way, while the board assumed that the 4 bytes were ordered in a different way. Okay, this was not pleasant, but the problem could be bypassed by reinterpreting the 32 bits in software.

软件更改后,芯片开始通过我的一些测试,但在一项更简单的测试中它们都失败了。更糟糕的是,即使是同一个芯片,每次测试运行时的失败情况也不同。因此,问题可能不在于 VLSI Technology 的制造测试设置。芯片内部发生了一些奇怪的事情。我尝试了各种假设。由于芯片在预制造模拟期间工作,因此最可能的原因是某种电气问题。问题是什么类型的电气问题。在其中一项试验中,芯片的行为似乎存在某种模式敏感性。该芯片的行为似乎有所不同,具体取决于最后读取的内容。为了测试是否确实存在特定于模式的灵敏度,我尝试在从芯片执行失败的读取操作之前将最后一个读数设置为 32 位零。该芯片现在通过了之前未能通过的测试。然后,我添加了一个读取操作,该操作似乎在每次正常读取操作之前都能可靠地读出零,并且该芯片通过了所有修改后的测试。我将这一发现告知了 VLSI Technology,并向他们发送了修改后的制造测试向量,以便我们能够尽快对所有进来的芯片进行工厂预筛选。

After the software change, the chips started to pass some of my tests, but they all failed on one of the simpler tests. Worse, even the same chip failed differently from one test run to another. So, the problem was probably not with VLSI Technology’s manufacturing test setup. Something peculiar was going on inside the chip. I tried out all sorts of hypotheses. Since the chip worked during the pre-fabrication simulation, the most likely cause had to be some sort of electrical problem. The question was what type of electrical problem. In one of the trial tests, there seemed to be some pattern sensitivity in the chip’s behavior. The chip appeared to behave differently depending on what was last read out from it. To test out whether there was indeed a pattern-specific sensitivity, I tried setting the last readout to be a 32-bit zero before doing a failed read operation from the chip. The chip now passed the test that it had previously failed. I then added a read operation that appeared to read out zeros reliably before every normal read operation, and the chip passed all the modified tests. I informed VLSI Technology about the finding and sent them the revised manufacturing test vectors, so that we could get factory prescreening on all the incoming chips as soon as possible.

仍然存在导致芯片奇怪行为的问题。我尝试设置额外的读取操作来读出 32 位的 1,芯片再次工作。我现在强烈怀疑问题的原因是芯片上 32 位输出总线之间的某种“交叉耦合”。当总线线路之一例如从1切换到0时,由于切换线路和相邻线路之间通过电容6的交叉耦合,相邻线路可能经历寄生电压降。在较旧的芯片技术中,交叉耦合效应并不是太重要,但随着技术的进步,导线间距变得更小,导线之间的电容达到了交叉耦合不可忽视的程度。

There was still the problem of what was causing the chips’ strange behavior. I tried setting the extra read operation to read out 32 bits of 1s, and again the chip worked. I now strongly suspected that the cause of the problem was some sort of “cross coupling” among the 32-bit output bus wires on the chip. When one of the bus wires switches, say, from 1 to 0, the adjacent wires could experience a spurious voltage drop due to cross coupling through the capacitances 6, between the switching wire and the adjacent wires. The cross coupling effects were not too important in older chip technologies, but with the more advanced technologies the wire spacing was much smaller, and the capacitances between the wires went up to the point that cross coupling could not be ignored.

VLSI Technology 的工程师一开始并不相信我的交叉耦合理论,但在与我讨论了输出总线线的电容值后,他同意这是合理的。

The engineer from VLSI Technology did not believe my cross coupling theory at first, but after going over the capacitance values for the output bus wires with me, he agreed that it was plausible.

如何避免交叉耦合问题?我们可以将总线线间隔得更远,以减少线间耦合电容。或者,可以在总线线之间引入屏蔽线。可能还有其他一些简单的解决方案。VLSI Technology 的软件没有提供任何现成的答案。我可以使用的一种解决方案是在读取输出值之前给输出总线更多的时间来稳定。当然,这需要修改电路设计并重新制造芯片。

What could be done to avoid the cross coupling problem? We could space the bus wires further apart to reduce the inter-wire coupling capacitance. Alternatively, shielding wires could be introduced between the bus wires. There were probably some other simple solutions. The software from VLSI Technology did not provide any ready answer. One solution at my disposal was to give the output bus more time to settle before reading the output value. Of course, that would require a modification to the circuit design, and re-fabricating the chips.

不管怎样,我感觉所有的芯片问题都得到了解决,于是我向团队的其他成员发送了一封电子邮件,总结了我的发现。第二天早上,我收到了乔·霍恩的消息。Joe继承Deep Thought II代码时,添加了大量用于程序自测试的代码。使用自测试代码,Joe 发现即使在添加预读操作后,芯片仍存在可重复的问题。经过一个漫长的夜晚,乔仔细地将问题归咎于与路过的国际象棋动作有关的问题。过路规则是我最不喜欢的国际象棋规则之一。我的半年多时间都浪费在解决一台又一台国际象棋机中与过路人有关的问题上。我花了半天时间才明​​白问题的确切本质。我犯了一个微妙的逻辑设计错误。当黑方采取过路棋时,筹码在某些条件下会出现异常行为。我能想到的最简单的临时解决方案是同时使用两个国际象棋筹码,其中一个筹码具有常规棋盘,另一个筹码将棋盘翻转 180 度,白棋变成黑棋,反之亦然。通过比较这两个芯片,可以即时定位问题并通过软件修复。添加预读操作已经使国际象棋程序变得更加复杂。新的翻板操作,加上预读操作,大概代码量增加了十倍。在我登上飞往北京的飞机之前,乔还有大约一周的时间来修复软件并查明是否还有其他问题。默里,他创造了软件对于新硬件的评估功能,也突然有了一些新的任务需要执行。同时使用两个芯片会带来一些复杂情况,必须在评估软件中解决。

Anyway, feeling that all the chip problems were resolved, I sent an e-mail to the rest of the team summarizing my findings. The next morning, I received a message from Joe Hoane. When Joe inherited the Deep Thought II code, he added a massive amount of code for program self-testing. Using the self-testing code, Joe found a repeatable problem with the chip even after adding in the pre-read operation. After a long night, Joe meticulously pinned down the problem to something related to the en passant chess moves. The en passant rule is one of my least favorite rules in chess. Over half a year of my life was wasted fixing problems related to en passant in one chess machine after another. It took me half a day to understand the precise nature of the problem. I had made a subtle logic design error. The chip misbehaved under certain conditions when Black had an en passant move. The easiest temporary solution that I could think of was to use two chess chips simultaneously, one of the chips having the regular board, and the other having the board flipped by 180 degrees with White pieces becoming Black pieces, and vice versa. By comparing the two chips, the problems could be located on the fly and fixed in software. Adding the pre-read operation had already made the chess program far more complicated. The new board flipping operation, combined with the pre-read operation, probably increased the amount of code tenfold. Joe had about one week left to fix the software and find out whether there was any other showstopper before I was to hop on the plane to Beijing. Murray, who was creating the software for the new evaluation function hardware, also suddenly had some new tasks to perform. Using two chips simultaneously introduced several complications that had to be resolved in the evaluation software.

乔发现了一些其他问题,但没有发现芯片出现新的主要问题。我很快对芯片设计进行了更改,并向 VLSI Technology 的工程师提供了新的网表。我们只有足够的时间为 1996 年 2 月的比赛获得一批新的筹码。

Joe uncovered some additional problems but no new major one with the chip. I quickly made changes to the chip design and gave the engineer at VLSI Technology the new netlist. We had just enough time to get a new batch of chips for the match in February 1996.

乔整整一周工作了很长时间,而新程序仍然不太有效。我们制作了一盘包含该软件快照的磁带。9月18日,我带着胶带和一块印有两个棋子的印刷电路板登上了飞机。乔还有大约 64 小时的时间来创建一个使用新芯片下棋的程序。穆雷留下来以防乔需要任何帮助。CJ Tan也参加了这次旅行,以帮助顺利解决北京的事情。

Joe worked very long hours during the entire week, and the new program was still not quite working. We made a tape containing a snapshot of the software. On 18 September, I got on the plane with the tape and a printed circuit board with two chess chips on it. Joe had about sixty-four hours left to create a program that used the new chips to play chess. Murray stayed behind in case Joe needed any help. C J Tan took the trip too, to help smooth things over in Beijing.

飞机旅行很顺利。抵达后的第二天早上,我向乔询问了情况。新计划仍在制定中,但他有一些事情要我检查。白天,我和 CJ 参观了 IBM 北京研究实验室,进行了初步设置。我尝试连接到美国,但电话线只是断断续续地工作,而且无法检索任何内容。北京实验室位于一个新开发区,几年前这里还是农田。事实上,我去的时候附近还有农田。显然,北京经济的快速发展对电话线路容量造成了严重的压力,而城市边远地区的服务也因此受到影响。

The plane trip was uneventful. I checked with Joe the morning after my arrival; the new program was still being worked on, but he had some things that he wanted me to look over. C J and I visited the IBM Beijing Research Lab during the day to do the preliminary setup. I tried to connect to the United States but the phone line only worked intermittently, and it was impossible to retrieve anything. The Beijing Lab was in a newly developed area that only a few years back was farmland. In fact there was still farmland nearby at the time of my visit. Apparently, the rapid economic development in Beijing had put serious strain on the phone line capacity, and the service in the outlying areas of the city suffered as a result.

回到酒店后,我向乔询问了新计划的进展情况。比赛原定于第二天下午举行。Joe 的东西几乎成功了,他相信他能够在比赛前通过 IBM 内部网络及时向我发送一个比赛程序。第二天一早,IBM中国总部的员工带着一张软盘来到了比赛现场。距离比赛还剩四个小时。将软盘上的程序加载到工作站硬盘上,我做了一些简单的测试。是的,该程序似乎正在运行。乔付出了英勇的努力。好吧,除了等待和祈祷之外,没有什么可做的。

After I got back to the hotel, I checked with Joe on the status of the new program. The match was to take place the next afternoon. Joe had something that almost worked, and he believed that he would be able to send a playing program to me in time for the match, via the IBM internal network. The next morning, an employee from IBM China Headquarters brought a floppy disk to the playing site. There were four hours left to the match. The program on the floppy disk was loaded onto the workstation hard disk, and I did some simple tests. Yes, the program appeared to be working. Joe had put in a heroic effort. Well, there was not much to do but wait and pray.

谢军是一位二十岁出头的娇小姑娘,在获得中国政府资助之前曾下过中国象棋“西方”国际象棋的训练。随后,她在世界国际象棋舞台上崭露头角,并赢得了女子世界国际象棋锦标赛的冠军,令国际象棋界感到惊讶,成为第一位获得此殊荣的亚洲人。由于我们的比赛是一场友谊赛,谢军必须选择时间控制。我没有关于使用什么时间控制的记录,但游戏大概是二十分钟,每一步之后还有二十秒。不保留游戏分数,这让谢军有机会尝试新的开局思路,而不必担心泄露她的秘密。我们事先约定要玩两场比赛。观众坐在附近的一个房间里,通过闭路视频连接与比赛室相连。谢军教练为观众进行跑动解说。

Xie Jun, a petite young lady in her early twenties, used to play Chinese chess before she received Chinese government sponsored training for “western” chess. She then burst onto the world chess scene, and surprised the chess establishment by winning the Women’s World Chess Championship, becoming the first Asian to do so. As our match was a friendly one, Xie Jun got to choose the time control. I don’t have a record of what time control was used, but it was probably twenty minutes for the game, with an additional twenty seconds after each move. No game score was kept which gave Xie Jun an opportunity to try out new opening ideas without worrying about revealing her secrets. We agreed beforehand to play two games. The audience sat in a nearby room linked to the playing room via a closed circuit video hookup. Xie Jun’s coach gave a running commentary to the audience.

第一局,谢珺显然是尝试了一个新的开局。这个长达 4 小时的节目在开场后很快就陷入了麻烦。一招,谢珺放慢了脚步,审视着自己的位置。此时新节目的成绩下降幅度较大。谢钧有一招杀招,她却没有看到,还下出了本该足以获胜的一招。然而,有趣的是,她在击杀的那一刻放慢了速度。又进行了几次操作后,新程序崩溃了。布莱克有可能以某种方式使程序崩溃。当我尝试重新启动程序时,我不小心删除了它保留的内部乐谱文件。谢军大方地答应玩新游戏。玩的是完全相同的游戏。谢钧在关键位置再次放慢了速度。她一定是感觉到自己还有更好的棋法可以下。由于时间控制太短,她再次错过了杀招,但她的这一招已经足够好了。程序在同一点再次崩溃7。这次我设法重新启动了程序。我们玩了一会儿,然后程序又以类似的方式崩溃了。由于节目的位置变得非常糟糕,我们在没有尝试重新开始的情况下就丢掉了第一场比赛。

In the first game, Xie Jun apparently tried out one of her new openings. The 4-hour-old program got into trouble pretty quickly after the opening phase. On one of her moves, Xie Jun slowed down and studied the position. At this point the new program had a big score drop. There was a killer move for Xie Jun but she did not see it, and played a move that should have been good enough to win. It was, however, interesting to observe that she slowed down at the precise moment when there was a kill. After a few more moves, the new program crashed. There was an en passant possibility for Black that somehow crashed the program. When I tried to restart the program, I accidentally wiped out the internal score file that it kept. Xie Jun graciously agreed to play a new game. Precisely the same game got played. Once again, Xie Jun slowed down at the critical position. She must have sensed that there was a better move that she could play. With the short time control, she again missed the killer move, but her move was good enough. The program crashed again at the same point7. This time I somehow managed to get the program restarted. We played on a little while, then the program crashed again in a similar fashion. Since the program’s position was getting pretty bad, we conceded the first game without trying to restart.

第一局比赛结束后,谢军提到,她一定是在关键位置错过了一个杀手锏。我证实了她的怀疑,并告诉她程序的计算结果。短暂的休息后,我们开始了第二场比赛。这次新程序运行时没有出现任何崩溃。谢军陷入战术困境,并在职位无望后立即辞职。看到她没有继续比赛,我松了口气。如果游戏持续更长时间,程序很可能会再次崩溃。

After the first game, Xie Jun mentioned that she must have missed a killer move in the critical position. I confirmed her suspicion and told her what the program had calculated it to be. After a short break, we started the second game. This time the new program played without any crashes. Xie Jun fell into a tactical bind, and resigned as soon as the position became hopeless. I was relieved that she did not play on. The program could very well crash again if the game went on much longer.

所以北京的比赛是一对一的平局,我们和卡斯帕罗夫的比赛还没有结束。

So the Beijing match was tied at one to one, and our match date with Garry Kasparov was still on.

“没有计算机能打败我”

“No Computer Will Ever Beat Me”

当我在卡内基梅隆大学时,计算机科学系的研究生组织操作一台可乐机,以每瓶 35 美分的价格出售廉价的瓶装可乐。这些收益是该部门许多 TG 政党的主要资金来源。我在卡内基梅隆大学呆了一两年左右,一些研究生和工程人员聚在一起,决定为可乐机制作一个电子界面。然后,人们就可以在办公室范围内通过电子方式“手指”它,并查明机器中是否遗留有任何冷瓶。事实上,您可以从加利福尼亚州帕洛阿尔托等地找到宾夕法尼亚州匹兹堡的可口可乐机器,尽管这种功能可能没有什么用处。

When I was at Carnegie Mellon, the graduate student organization in the Computer Science Department operated a Coke machine that dispensed cheap bottled Coke for thirty-five cents a bottle. The proceeds were the main funding source for many of the TG parties in the Department. About a year or two into my stay at Carnegie Mellon, some graduate students and engineering staff got together and decided to make an electronic interface for the Coke machine. It then became possible to “finger” it electronically from the confines of one’s office, and find out whether any cold bottles were left in the machine. In fact, you could finger the Coke machine in Pittsburgh, Pennsylvania from, say, Palo Alto, California, although this capability probably offered very little utility.

可乐机位于我所在大楼的三楼。我的办公室在八楼,所以我摸索了很多次,直到我找到了替代供应——我所在楼层的一个较大的办公室有一台大冰箱,几个学生将其用作辅助可乐机,作为交换为了特权,会帮忙从三楼的储藏室装冰箱。

The Coke machine was located on the third floor of my building. My office was on the eighth floor so I fingered it a lot, until I found an alternative supply—one of the bigger offices on my floor had a large refrigerator that was used as an auxiliary Coke machine by a few students, who, in exchange for the privilege, would help load up the refrigerator from the storage room on the 3rd floor.

杰伊·西佩尔斯坦就是这间大办公室的主人。杰伊在耶鲁大学获得了学士学位,在耶鲁大学期间,他经常在耶鲁国际象棋俱乐部闲逛。有一天,杰伊向我提到,有一位耶鲁大学的学生,一位国际大师,他会在国际象棋俱乐部同时表演并击败所有人。他说有人问国际大师是否认为任何计算机都能打败他。国际大师回答说:“没有计算机能打败我。” 当他吹嘘的时候,他可能有充分的理由沾沾自喜。几年前,他曾在电视节目中在闪电战游戏中屠杀克雷·闪电战。他也是一位后起之秀。国际大师从耶鲁大学毕业后的第二年,他就成为了特级大师。这位新大师的名字叫乔尔·本杰明。

Jay Sipelstein was an occupant of this big office. Jay had received his Bachelor degree from Yale University and while he was there, he used to hang around the Yale Chess Club. One day, Jay mentioned to me that there was a Yale student—an International Master—who would give simultaneous exhibitions in the chess club and beat everybody. He said that someone had asked the International Master whether he thought any computer would ever beat him. The International Master answered, “No computer will ever beat me.” At the time he made the boast, he probably had good reasons to be smug. A few years before, he had been on a television show slaughtering Cray Blitz in blitz games. He was also an up-and-comer. The year after the International Master graduated from Yale, he became a Grandmaster. The name of this new Grandmaster was Joel Benjamin.


在北京对阵谢军的平局中保住了比赛加里·卡斯帕罗夫还活着。谢军是特级大师,但我们的友谊赛是在缩短的时间控制下进行的,我们想在更严峻的条件下进行一些测试赛。

The tied match against Xie Jun in Beijing kept the match against Garry Kasparov alive. Xie Jun was a Grandmaster, but our friendly exhibition was played at a shortened time control, and we wanted to play some test matches under more serious conditions.

我们没有深蓝的全部能力,它比我们拥有的机器快大约三十倍,所以在常规锦标赛时间控制下与大师们进行平等的比赛将是一个令人满意的结果。我们计划在实验室里与三位领先的美国特级大师进行比赛,每场两场比赛。特级大师们会受到激励,尝试赢得比赛,如果失败,则尝试打平。我们还利用测试赛来筛选特级大师,以帮助为与加里·卡斯帕罗夫的比赛做开场准备,但没有一个特级大师知道职位空缺。我们一直都知道在决赛中我们需要一位大师。当然,香港事件也已经说明了这一点。

We did not have the full power of Deep Blue, which would be about thirty times faster than the machine we had, so playing at an even level with Grandmasters at a regulation tournament time control would be a satisfactory result. We scheduled to play in our lab, two games each, against three leading US Grandmasters. The Grandmasters were offered incentives to try to win the match or, failing that, to try to draw the match. We were also using the test matches to screen for a Grandmaster to help with the opening preparation for the match against Garry Kasparov, but none of the Grandmasters knew about the job opening. We always knew that we would need a Grandmaster on the team in the final match. Of course, the Hong Kong incident had also made that crystal clear.

深蓝Jr,缩小版的机器,赢了一场比赛,平了另一场,输了一场。

Deep Blue Jr, the scaled down machine, won one of the matches, drew another, and lost one.

获胜的大师不是别人,正是乔尔·本杰明。Deep Blue Jr 正处于开发的早期阶段,因此输掉一场比赛并不令人意外。然而,乔尔以 2-0 赢得比赛的方式却令人大吃一惊。

The winning Grandmaster was none other than Joel Benjamin. Deep Blue Jr was in the very early stages of its development, so losing one of the matches was not a big surprise. The way Joel won the match 2–0, however, was a big surprise.

与乔尔的第一场比赛是一场典型的电脑比赛。深蓝Jr误判了终局位置,而乔尔的位置判断却更胜一筹。我们知道深蓝Jr的评估功能还有很大的改进空间,所以没有什么值得担心的。

The first game against Joel was a typical computer affair. Deep Blue Jr misjudged the endgame position, and Joel’s positional judgment proved superior. We knew that there was lots of room for improvement in Deep Blue Jr’s evaluation function, so there was nothing to be alarmed about.

第二场比赛是一场漫长的比赛。下午开始,深夜结束。当进入最后阶段时,很明显乔尔更好,但位置可能已经平局。提供的奖励是为了赢得或平局比赛,所以乔尔要做的实际事情就是提供平局,赢得比赛,接受支票,并说谢谢。乔尔继续追问。默里和我现在已经饿了。有一段时间,乔尔似乎没有取得任何进展。但不知怎么的,他却慢慢地、耐心地让小深蓝犯了一个又一个小错误,直到完全掌控了局面。然后他向深蓝施压,要求他最终放弃材料,不久之后我们就退出了这个项目。

The second game was a long one. It started in the afternoon and finished late at night. When the endgame phase was reached, it was clear that Joel was better but the position was probably drawn. The incentives offered were for winning or drawing the match, so the practical thing for Joel to do was to offer a draw, win the match, take the cheque, and say thank you. Joel pressed on. Murray and I were getting hungry by now. For a short while, it appeared that Joel was not making progress. But somehow he slowly and patiently tricked Deep Blue Jr into making one slight mistake after another, until he got complete control of the situation. He then pressured Deep Blue Jr to finally give up material, and shortly afterwards we resigned for the program.

送乔尔走出大楼后,默里和我沉默了一会儿。前面还有很多工作要做,但我们都知道我们已经找到了负责开幕准备的人对于深蓝来说。乔尔还不知道,但我们已经找到了最后一块拼图。问题是我们是否有足够的时间来创造出可以在 1996 年 2 月击败加里·卡斯帕罗夫的东西。

After walking Joel out of the building, Murray and I were silent for a while. There was a lot of work ahead, but we also both knew that we had found the person to do the opening preparation for Deep Blue. Joel did not know it yet, but we had found the last piece of the puzzle. The question was whether we had enough time to create something that could beat Garry Kasparov in February 1996.

    当安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 读到这本书的初稿时,他对我在事故发生前的想法感到惊讶。对我来说,这个项目就是我这段时间的生活。另外,我刚从与该项目相关的出差回来,这个项目就在我的脑海里。当时,该项目的生存完全取决于新的国际象棋芯片是否有效。由于我是该芯片的唯一设计者,如果我没有在事故中幸存,该项目将不复存在。

    When Andreas Nowatzyk read an early draft of this book, he was surprised at what I was thinking right before the accident. To me, the project was my life during this period. Also, I was returning from a business trip related to the project, and the project was on my mind. At that point in time, the survival of the project relied entirely on whether the new chess chip worked. Since I was the sole designer of the chip, if I had not survived the accident, the project would cease to exist.

    事故发生后一年多的时间里,我的视力仍然不时出现问题。

    For over a year after the accident, I continued to have problems with my eyesight from time to time.

    如果我们能够使用更好的技术,国际象棋筹码的速度大约会提高两倍。芯片面积应减少约四倍,这将显着提高芯片产量。

    The chess chips would have been about twice as fast if we had had access to the better technology. The die area should reduce by about a factor of four, which would improve the chip yield quite dramatically.

    该事件将在下一章中描述。我对第一场深蓝比赛的第一次讨论是在活动结束时进行的。

    This event will be described in the next chapter. The first discussions I had about the first Deep Blue match took place at the end of the event.

    欧姆表是用于测量电阻的电表。

    An ohmmeter is an electric meter used to measure electrical resistance.

    这并不严格正确,但在高频开关期间,两根导线之间的高电容可以被视为一个小(虚数)电阻。因此,一根导线的电压变化将通过电容“传导”到另一根导线,尤其是当电容较高时。

    This is not strictly correct, but a high capacitance between two wires can be viewed as a small (imaginary) resistance during high frequency switching. So the voltage changes from one wire will be “conducted” through the capacitance to the other wire, especially when the capacitance is high.

    我们从未发现导致该版本程序崩溃的错误。当我从中国回来时,乔已经让程序正常运行了。

    We never found the bug that caused the crashes in this version of the program. By the time that I got back from China, Joe had the program working properly.

 

 

第10章

CHAPTER 10

活生生的珠穆朗玛峰

A Living Mount Everest


在海滨度假胜地

At a Seaside Resort

早在香港赛事之前,我就意识到 ACM 有兴趣在不寻常的情况下举办一场卡斯帕罗夫与深蓝的比赛。

I became aware of the ACM’s interest in hosting a Kasparov vs. Deep Blue match under unusual circumstances, well before the Hong Kong event.

当我在研究深蓝国际象棋芯片时,我认为参加计算机国际象棋比赛很麻烦。即使我没有参与这次活动的准备工作,我也清楚地知道,几周内我将无法在国际象棋芯片上做任何认真的工作——我无法阻止自己关注我们在活动中的进展。

While I was working on the Deep Blue chess chip, I considered attending computer chess events a nuisance. Even if I had not been involved with the preparation for the event, I knew perfectly well that I would not be able to do any serious work on the chess chip for a few weeks—I could not stop myself from following our progress in the event.

当 1994 年 ACM 计算机国际象棋锦标赛到来时,我反对参加。国际象棋芯片设计工作已进入关键的最后阶段,考虑到整个项目都取决于芯片的工作,在一场重要的比赛上浪费时间似乎有些鲁莽。ACM 赛事的组织者、CJ Tan 的好友 Monty Newborn 说服他参加 Deep Thought II 锦标赛,但我被否决了。

When the 1994 ACM Computer Chess Championship came along, I was against going. The chess chip design work was entering the critical final phase, and given that the entire project hinged on the chip working, it seemed foolhardy to waste time on a tournament, albeit an important one. Monty Newborn, the organizer of the ACM event, and a good friend of C J Tan, persuaded him to enter Deep Thought II into the Championship, and I was overruled.

原来,穆雷和乔都已经在夺冠前后安排好了假期。乔也许能够参加锦标赛的最后部分,但肯定会错过早期部分。我是唯一一个可以参加ACM活动并操作该程序的人。

It turned out that both Murray and Joe had already scheduled their vacations around the time of the championship. Joe might be able to show up for the closing part of the championship, but would definitely miss the early part. I was the only person who could go to the ACM event and operate the program.

锦标赛在新泽西州南端的海滨度假小镇五月角举行,距离我们位于霍桑的实验室约四小时车程。由于我无法避免去的“义务”,所以我把这次旅行当作芯片设计工作的假期1。从这个意义上说,驾驶过程相当愉快——风景优美的路线、阳光、新鲜的空气和充足的海风。在酒店,我遇到了其他团队的几位成员。每个人都试图互相感受。今年谁会是危险的?来自麻省理工学院的明星苏格拉底吸引了很多关注,因为这将是自 Deep Thought II 以来第一个突破每秒 100 万个位置大关的程序。

The championship was held in Cape May, a seaside resort town at the southern tip of New Jersey, about a four-hour drive from our lab in Hawthorne. Since there was not much I could do to avoid the “obligation” to go, I treated the trip as a vacation from the chip design work1. In that sense, the drive was quite pleasant—a scenic route, sun, fresh air, and lots of sea breeze. At the hotel, I bumped into several members of the other teams. Everyone was trying to feel each other out. Who would be dangerous this year? Star Socrates from MIT was attracting a lot of attention, as it would be the first program since Deep Thought II to cross over the one million positions per second mark.

比赛第一天,天气突然暴风雨。我不知道的是,霍桑的风暴要严重得多。我们第一轮的对手是扎尔科夫,他是一位令人尊敬的对手。比赛之前,穆雷一直在尝试国王弃兵开局。国王弃兵是一种暴力开局,可以追溯到 19 世纪国际象棋的浪漫时代。如今在顶级国际象棋赛事中很少见到这种情况。对于白方来说,这也可能不是一个良好的开局。默里在“深思二号”的帮助下发现了白棋在国王弃兵中的一些开局创新,并想尝试一下。然而扎尔科夫并没有合作。它所走的路线通向穆雷没有检查过的位置,这对于黑棋来说是一个很好的,甚至可能是一个获胜的位置。深思二号对自己所处的局面不太满意,但黑方并没有明显的杀戮。所有其他来到我们桌子旁的团队都认为我们迷路了。说实话,我并不太担心,但我应该担心。无知是福。如果他在场,穆雷就会自责。我见过“深思二号”使用国王弃兵出色地获胜,尽管它通常不喜欢白方在开局时的位置。我只是不知道,如果扎科夫知道如何正确下棋,棋盘上的位置不会给我们反击的机会。在扎尔科夫无意中的帮助下,《深思二号》抓住了足够多的机会,最终《扎尔科夫》的作者约翰·斯坦巴克在职位无望后就为扎尔科夫辞职了。当时我和他都不知道,如果比赛再持续半个小时,《深思二》就不得不认输了。

On the first day of the tournament, the weather turned stormy. What I did not know was that the storm was much worse in Hawthorne. Our first round meeting was with Zarkov, a respectable opponent. Prior to the tournament, Murray had been experimenting with the King’s Gambit opening. The King’s Gambit is a violent opening dating back to the romantic age of chess in the 19th century. It is rarely seen these days at top-level chess events. It is also probably not a sound opening for White. Murray had found some opening innovations for White in the King’s Gambit with the help of Deep Thought II, and wanted to try them out. Zarkov, however, was not cooperating. It played a line that led to a position which Murray had not examined and it was a good, possibly even a winning, position for Black. Deep Thought II was not too happy about the situation it found itself in, but there was no obvious kill for Black. All the other teams who came by our table thought that we were lost. To tell the truth, I was not too concerned, but I should have been. Ignorance is bliss. Had he been present, Murray would have been blaming himself. I had seen Deep Thought II win brilliantly using the King’s Gambit, even though it usually did not like White’s position out of the opening. I just did not know that the position on the chessboard gave us no counter chance, if Zarkov had known how to play it correctly. With Zarkov’s unwitting help, Deep Thought II seized enough opportunities, and eventually John Stanback, the author of Zarkov, resigned for Zarkov as soon as the position became hopeless. Neither he nor I knew at the time that if the game had lasted for another half-hour, Deep Thought II would have had to forfeit.

当我们对阵扎尔科夫的比赛正在进行时,霍桑突然下起了倾盆大雨。一小时内降雨量超过两英寸,积水导致变压器短路实验室的电源。在与扎科夫比赛的最后半小时,Deep Thought II 使用我们的 UPS(不间断电源)备用电池供电。电池只能持续一个小时左右,到第二轮就耗尽了电量。我发现我根本无法连接到Deep Thought II,所以我向赛事总监请求暂停以调查到底发生了什么。我给我们的支持工程师2杰里·布罗迪 (Jerry Brody) 打电话,他是团队中唯一不在城外的人,并请他进入实验室找出问题所在。二十分钟后,杰瑞告诉了我这个坏消息。实验室里没有电,也不知道什么时候能恢复。经过对手的同意,赛事总监给了我们更多的时间。在计算机国际象棋比赛中,通常会规定无限的通信暂停。规则中从未明确涵盖电源故障,尽管可以认为它等同于通信故障。到晚上十一点,我们还是没能把机器拿回来。我们的对手被问到比赛是否可以改到第二天进行,他回答说。他拒绝了,我们没有下任何一步就放弃了比赛。机器在凌晨一点左右启动并运行,但为时已晚。(十八个月后,在卡斯帕罗夫与深蓝的第一场比赛中,我们向电力公司通报了我们的电力需求。公司在看到比赛的头条新闻后,做出了特殊安排,以确保电力的持续供应他们当然不希望因为自己的电源故障导致深蓝输掉一场比赛而在全世界范围内声名狼藉!)

While our game against Zarkov was going on, a sudden downpour took place in Hawthorne. Over two inches of rain fell within an hour, and the buildup of water shorted the transformer providing power to the lab. For the last half-hour of the game against Zarkov, Deep Thought II was using the battery power of our UPS (Uninterruptible Power Supply) backup. The battery could last only an hour or so, and it ran out of power by the second round. I found that I could not connect to Deep Thought II at all, so I asked for a time-out from the tournament director to investigate what was going on. I phoned Jerry Brody, our support engineer2, and the only person in the group not out of town, and asked him to get into the lab to find out what went wrong. Twenty minutes later, Jerry gave me the bad news. There was no power in the lab, and there was no estimate about when it would be back. With the consent of our opponent, the tournament director granted us more time. In computer chess events, there is usually a provision for unlimited communication time-out. Power failure was never explicitly covered in the rules, although an argument could be made that it was equivalent to a communication failure. By eleven pm, we still could not get the machine back. Our opponent was asked whether the game could be rescheduled to the next day; he declined and we forfeited the game without playing a single move. The machine was up and running at about one am, but it was too late. (Eighteen months later, during the first Kasparov vs. Deep Blue match, we informed the electric company about our power requirements. After seeing the front-page news that the match was making, the company made special arrangements to ensure a continuous supply of electric power to IBM Research Lab. They certainly did not like the prospect of getting worldwide notoriety for causing Deep Blue to lose a game due to a power failure on their part!)

回到锦标赛。该赛事为五轮瑞士赛。出于所有实际目的,因弃权而输掉一轮几乎消除了我们赢得锦标赛的机会。我们能做的最好的就是 5 分中的 4 分。根据过去的经验,这通常不足以赢得像这样的比赛。我感到很痛苦。

Back to the tournament. The event was a five-round Swiss. For all practical purposes, losing one round by forfeit pretty much eliminated our chance of winning the tournament. The best we could do would be 4 points out of 5. Based on past experience, this was usually insufficient to win a tournament like this one. I was feeling miserable.

第三轮我们的对手是WCess,他们打得很激烈,但Deep Thought II轻松获胜。第四轮的对手是明星苏格拉底,我们将面对他一年后再次来到香港。在游戏过程中,麻省理工学院的团队非常高兴地发现 Star Socrates 报告的搜索深度比 Deep Thought II “更大”。在本阶段的比赛中,苏格拉底之星是唯一一支获得满分的队伍。还有怀特,所以麻省理工队一定觉得他们很有机会赢得比赛。与一年后的香港比赛一样,深思II相对顺利地赢得了比赛,尽管两场比赛都有焦灼的时刻。毕竟,任何每秒搜索超过一百万个位置的东西都不能掉以轻心,而明星苏格拉底绝对是一个竞争者。

Our third round opponent was WChess which put up a stiff fight, but Deep Thought II won without too much trouble. The fourth round opponent was Star Socrates, whom we would face again a year later in Hong Kong. During the game, the MIT team was very pleased to observe that Star Socrates was reporting “greater” search depth than Deep Thought II. At this stage of the tournament, Star Socrates was the only team with a perfect score. It also had White, so the MIT team must have felt that they had a great chance to win the event. Just as in Hong Kong a year later, Deep Thought II won the game relatively smoothly, even though there were anxious moments in both games. After all, anything that searches over a million positions per second is not something to be taken lightly, and Star Socrates was definitely a contender.

第五轮配对变得非常有趣。唯一保持不败的程序是 Mchess,我们在第二轮比赛中输给了他。米凯斯只丢了一场平局。通常,在瑞士锦标赛中,您不会在同一锦标赛中与任何对手交手两次。但 Mchess 和 Deep Thought II 从未真正对战过,而且 Deep Thought II 已经玩过了所有其他顶级程序,除了 Mchess 之外。于是 Deep Thought II 再次下起了 Mchess,这次是真的,并且很快就赢得了比赛。比赛现在的问题是,明星苏格拉底能否赢得最后一场对阵扎科夫的比赛,如果是的话,它将并列第一。明星苏格拉底输了,深思二号干净利落地赢得了冠军,尽管第二轮就被淘汰了。

The fifth round pairing became very interesting. The only program that remained undefeated was Mchess, the second round opponent against whom we had forfeited. Mchess had conceded only one draw. Normally, in a Swiss tournament, you would not play any of your opponents twice in the same tournament. But Mchess and Deep Thought II never did really play each other, and Deep Thought II had already played all the other top programs, with the sole exception of Mchess. So Deep Thought II played Mchess again, this time for real, and won the game fairly quickly. The tournament now became a question of whether Star Socrates could win its last game against Zarkov, in which case it would tie for first. Star Socrates lost, and Deep Thought II won the title cleanly, despite the second round forfeit.

第二天早上,我在酒店吃了一顿丰盛的早餐。阳光明媚。海滩看起来很诱人,“深思二号”完成了一项不可能的壮举。我的心情非常好。比赛组织者蒙蒂·纽伯恩 (Monty Newborn) 问他和他的朋友是否可以加入我。蒙蒂介绍他的朋友是费城附近天普大学的教授弗兰克·弗里德曼。Frank 也是 Monty 在 ACM 组织中的老板。蒙蒂和弗兰克问我新的国际象棋芯片什么时候能准备好,我告诉他们应该会在明年的某个时候上市。蒙蒂随后提出了真正的话题。他问道:“IBM 有兴趣在 1996 年 ACM 年会上与加里·卡斯帕罗夫进行一场比赛吗?” 这变得有趣了。我回答说:“我不代表 IBM 发言,但我想是这样。” 然后我补充道:“这场比赛会很贵。” 蒙蒂表示,他预计这场比赛的奖金约为30万至40万美元。我告诉他我相信奖金一定要高得多,因为这场比赛比常规的世界象棋比赛具有更大的历史意义冠军赛。上届世界国际象棋锦标赛的奖金远远超过一百万美元。我告诉蒙蒂和弗兰克:“如果加里只接受低于 100 万美元的价格,我会感到非常惊讶。” 蒙蒂有自己的一些想法,“好吧,如果我们只打六场比赛,那么我们就可以减少奖金。” “也许”,我没有多少信心地回答。我们结束了谈话,但没有做出任何真正的承诺。ACM 将开始与 Garry 交谈并获取他的反馈。我认为他们让卡斯帕罗夫加入的几率为五分之一。

The next morning, I had a good breakfast in the hotel. The sun was shining. The beach looked enticing, and Deep Thought II had pulled off an impossible feat. I was in a very good mood. Monty Newborn, the tournament organizer, asked whether he and his friend could join me. Monty introduced his friend as Frank Friedman, a professor from Temple University near Philadelphia. Frank was also Monty’s boss in the ACM organization. Monty and Frank asked me when the new chess chip would be ready, and I told them that it should be available some time the following year. Monty then broached the real subject. He asked, “Would IBM be interested in playing a match with Garry Kasparov during the ACM Annual Conference in 1996?” This was getting interesting. I answered, “I don’t speak for IBM, but I would imagine so.” Then I added, “The match would be quite expensive.” Monty said that he expected the prize fund for the match to be something like $300,000 to $400,000. I told him I believed that the prize fund would have to be much higher, because the match would have greater historical importance than a regular World Chess Championship match. The last World Chess Championship match had a prize fund of well over one million dollars. I told Monty and Frank, “I would be very surprised if Garry settles for anything less than a million dollars.” Monty had some ideas of his own, “Well, if we only play, say, six games, then we could get the prize fund down.” “Perhaps”, I answered without much conviction. We concluded the conversation without making any real commitment. ACM would start talking with Garry and get his feedback. I placed their odds of getting Kasparov on board at 1 in 5.

ACM 的人证明我错了。加里最终同意在费城与“深蓝”进行六场比赛,总奖金为 50 万美元,其中获胜者获得 40 万美元,失败者获得 10 万美元。加里对赢得比赛非常有信心,因此他没有要求更多的钱就同意了。事实上,他希望这场比赛要么全有要么全无,失败者一无所获,但最终被说服放弃了这个想法。IBM 研究中心设立了奖金。深蓝公司获得的奖金将返还给 IBM,用于资助更多的研究,但不一定是计算机国际象棋方面的研究。ACM 将这场比赛作为其成立 50 周年庆典的一部分。1996 年也是第一台电子计算机 ENIAC 诞生 50 周年,ENIAC 在费城附近建造。因此,ACM 决定在费城拉开为期一年的庆祝活动,卡斯帕罗夫对阵深蓝的比赛(1996 年 2 月 10 日至 17 日)是议程上的第一项活动。比赛以每天一场比赛的速度进行,每两个比赛日休息一天。比赛日为2月10日、11日、13日、14日、16日和17日。

The ACM people proved me wrong. Garry eventually agreed to play a six-game match against Deep Blue in Philadelphia for a total prize fund of $500,000, with $400,000 to the winner, and $100,000 to the loser. Garry was so confident of winning the match that he agreed to it without demanding more money. In fact, he wanted the match to be all or nothing with the loser getting nothing but was eventually talked out of the idea. IBM Research put up the prize fund. Deep Blue’s share of the prize would go back to IBM to fund more research, not necessarily in computer chess. ACM was scheduling the match as part of its 50th anniversary celebration. 1996 was also the 50th anniversary for ENIAC, the first electronic computer, and ENIAC was built near Philadelphia. Therefore, ACM decided to kick off its year-long celebration in Philadelphia, and the Kasparov vs. Deep Blue match (10–17 February 1996) was the first event on the agenda. The match was to be played at the rate of one game per day, with one rest day following every two game days. The game days were 10, 11, 13, 14, 16 and 17 February.

费城的冬天

Winter in Philadelphia

1995年12月,费城比赛前两个月,我们收到了VLSI Technology公司的一小批经过修改的国际象棋芯片。这次芯片工作正常。不再需要进行预读操作。不再有en passant 的问题,也不再有芯片镜像。大师乔尔·本杰明同意成为深蓝的第二名。距离比赛剩下的时间太短了,他的主要职责就仅限于准备开局动作。由于大部分芯片仍在 VLSI 的测试和封装线上,深蓝尚未真正存在,乔尔几乎所有的开业准备工作都是在 Deep Blue Jr(深蓝的缩小版工作站版本)上完成的。

In December 1995, two months before the Philadelphia match, we received a small batch of the revised chess chips from VLSI Technology. The chips worked properly this time. There was no more need to do the pre-reading operation. There were no more problems with en passant, and no more chip mirroring. Grandmaster Joel Benjamin agreed to be Deep Blue’s match second. The time remaining until the match was so short that his main function was limited to preparing the opening moves. Since most of the chips were still in VLSI’s testing and packaging lines, Deep Blue did not really exist yet, and Joel did pretty much all his opening preparation on Deep Blue Jr, the scaled down workstation version of Deep Blue.

比赛前大约两周,剩余的筹码到达了。我们花了大半周的时间来测试它们,然后花了两天左右的时间将它们加载到 IBM RS/6000 SP 超级计算机(即深蓝的主机)中。赛前我们几乎没有足够的时间来测试所有的硬件,并在大机器上进行基本的软件测试。我们也没有得到我们想要的那么多棋子。

About two weeks before the match, the remainder of the chips arrived. It took the better half of a week to test them all out, and then two days or so to load them up into the IBM RS/6000 SP supercomputer, that was to be Deep Blue’s host. We had barely enough time before the match to test out all the hardware, and do the basic software testing on the big machine. We also did not have as many chess chips as we would have liked.

鉴于我们在香港的经验,我们决定在费城的比赛场地设置两台备用计算机,而主要的 RS/6000 SP 超级计算机留在纽约约克敦高地的实验室3。一台备份机器是运行 Deep Blue Jr 程序的常规 IBM RS/6000 工作站。另一台机器是一台小型四路 RS/6000 SP 超级计算机,有四个工作站节点。主机是一台三十六路4 RS/6000 SP超级计算机。理论上,对于 36 路机器,系统最多可以有 576 个国际象棋筹码一起工作,每个工作站节点有 16 个国际象棋筹码。由于可用棋子数量有限,主机上每个节点只有6个棋子,总共216个棋子。这批棋子每秒搜索约160万个位置,这第一颗深蓝的理论最大搜索速度约为每秒3亿个位置。观察到的搜索速度约为每秒 1 亿个位置。与并行运行搜索相关的软件开销相当高。如果我们再有六个月的时间来开发该软件,我们可能可以将机器的速度提高两倍。

Given our experience in Hong Kong, we decided to set up two backup machines at the playing site in Philadelphia, with the main RS/6000 SP supercomputer staying behind at the lab in Yorktown Heights, New York3. One backup machine was a regular IBM RS/6000 workstation running the Deep Blue Jr program. The other machine was a small four-way RS/6000 SP supercomputer, with four workstation nodes. The main machine was a thirty-six-way4 RS/6000 SP supercomputer. Theoretically, with a thirty-six-way machine, the system could have up to 576 chess chips working together, with sixteen chess chips per workstation node. Due to the limited number of chess chips available, each node on the main machine only had six chips, for an aggregate of 216 chess chips total. The chess chips in this batch searched about 1.6 million positions per second, and the theoretical maximum search speed of this first Deep Blue was about 300 million positions per second. The observed search speed was about 100 million positions per second. The software overhead associated with running the search in parallel was quite high. We probably could have speeded up the machine by up to a factor of two, if we had had another six months to work on the software.

国际象棋芯片对棋局的评估比任何其他国际象棋程序都要详细得多。几乎所有的评估函数项都是直接在芯片5上计算的。事实上,要在通用计算机上执行与 1996 年深蓝所做的相同的计算,将每秒至少需要一万亿条指令的通用系统6。这是有史以来最强大的国际象棋机器。在与费城的比赛之前,它也从未下过一场国际象棋比赛。

The chess chips evaluated the chess positions in far more detail than possibly any other chess program that had ever existed. Almost all the evaluation function terms were computed directly on chip5. As it was, to perform the same computation on a general purpose computer, as that done by the 1996 Deep Blue, would need at least a one trillion instructions per second general-purpose system 6. It was the most powerful chess machine ever built. It had also never played a single chess game going into the Philadelphia match.

费城距离很近,开车去那里比乘坐商业航班更有意义。1996年2月8日,穆雷、乔、杰里·布罗迪和我完成了最后一次赛前测试,开车前往费城,而特级大师乔尔·本杰明则乘坐火车。我不知道乔尔在旅途中的感受如何,但在我们的车里,气氛有点阴沉。我们已经走了这么远,但距离我们想要的目标还有很长的路要走。深蓝还是个两周大的婴儿。这是一个非常强大的婴儿,但仍然是一个婴儿。勒死赫拉女神派来的两条蛇的会是小赫拉克勒斯吗?还是我们派了一个无助的婴儿作为贡品来安抚海怪鲸鱼座,但没有珀尔修斯的帮助?我们担心会是后者。

Philadelphia was sufficiently close that driving there made more sense than taking a commercial flight. On 8 February 1996, Murray, Joe, Jerry Brody, and I finished the last pre-match testing and drove to Philadelphia while Grandmaster Joel Benjamin took the train. I don’t know how Joel felt during his trip but in our car, the atmosphere was a little bit somber. We had come so far, and yet we were still a long way from where we wanted to be. Deep Blue was still a two-week old baby. It was a very powerful baby, but a baby nonetheless. Would it be the baby Hercules that strangled the two serpents sent by the Goddess Hera? Or were we sending a helpless baby up as a tribute to placate the sea monster Cetus, but without the aid of Perseus? We were afraid that it would be the latter.

下午晚些时候,我们抵达费城万豪酒店。会议和比赛将在街对面的费城会议中心举行。一座封闭的步行桥将酒店直接连接到市中心。桥头会议中心的入口大厅装饰着几座高塔,让我想起了国际象棋的车。入住酒店时,迎接我们的是 ACM 公共关系总监 Terrie Phoenix 和 IBM Research 负责国际象棋项目新闻关系的公关人员 Marcy Holle。几个月前,杰里·普斯特退休后,马西接替了他。特里和玛西正在前往其他地方的路上,但他们通知我们第二天要举行新闻发布会。

We arrived at the Philadelphia Marriott Hotel in the late afternoon. The conference and the match were to be held in the Philadelphia Convention Center right across the street. An enclosed walking bridge connected the hotel directly to the center. The entrance hall at the convention center end of the bridge was decorated with several high towers that reminded me of chess rooks. While checking into the hotel, we were greeted by Terrie Phoenix, the ACM director of public relations, and Marcy Holle, the communications person from IBM Research in charge of press relations for the chess project. Marcy had replaced Jerry Present a few months back when he retired. Terrie and Marcy were on their way somewhere else, but they informed us about the press conference to be held the next day.

一小群 IBM 员工提前几天抵达,负责处理所有后勤、视频设置、Web 内容分发等事宜。我们与其中一些人共进晚餐。他们向我们讲述了他们遇到的一些惊喜。

A small contingent of IBMers had arrived several days earlier to deal with all the logistics, video setup, web content distribution, and so on. We met with some of them for dinner. They told us some of the surprises that they encountered.

比赛合同签订时,决定比赛在会展中心的大剧场进行,以满足想要直接观看比赛的观众。然而,在押金付清之前,负责安排的人就离开了ACM。等ACM赶回来的时候,剧院已经租出去了。一间演讲室被用作比赛室,毗邻的房间被保留为加里的更衣室。Matt Thoennes 是负责使整个安排顺利进行的 IBM 员工。报告厅里搭建了一个升高的舞台,并安装了多台摄像机为观众提供视频直播。空间足够容纳大约七八排观众席。当加里的母亲克拉拉·卡斯帕洛娃视察比赛场地时,她反对设置一排观众的想法。添加了一堵隔墙,只为 Garry 的随行人员、ACM 官员和 IBM 人员留下了一排座位。付费观众则留在解说室。至少从马特的角度来看,克拉拉的另一个有趣的要求是在加里的更衣室里安装一个 port-a-john(便携式厕所)。马特(Matt)向 ACM 的某人询问如何获得沐浴露,并惊讶地收到了各种沐浴露的目录,从最基本的到配有豪华设施的豪华行政车型。除了厕所外,还提供食物和饮料(包括三角巧克力,显然是加里的最爱之一)。

When the match contract was signed, it was decided that the match would be played in a large theater in the convention center, to accommodate spectators who wanted to see the match directly. However, the person in charge of making the arrangements left ACM before the deposit had been put down. By the time ACM got back to it, the theater had already been rented out. A lecture room was found to serve as the match room, with an adjoining room reserved as Garry’s dressing room. Matt Thoennes was the IBMer in charge of making the whole arrangement work. A raised stage was set up in the lecture room, and several video cameras were brought in to provide video feeds to the audience. There was enough space for about seven or eight rows of spectator seats. When Garry’s mother, Clara Kasparova, inspected the match site, she objected to the idea of having rows of spectators. A partitioning wall was added leaving just one row of seats for Garry’s entourage, ACM officials, and IBM personnel. The paying audience would stay in the commentary room. Another interesting request by Clara, at least from Matt’s point of view, was for a port-a-john (portable toilet) in Garry’s dressing room. Matt asked someone at the ACM how to get a port-a-john, and was surprised to receive a catalog of all sorts of them, from the most basic to fancy executive models with luxurious amenities. Besides the port-a-john, food and drink (including Toblerone chocolate, apparently one of Garry’s favorites) were provided.

我们抵达后的第二天早上,当我们在酒店餐厅吃早餐时,克拉拉和加里的教练尤里·多霍基安 (Yuri Dohokian) 坐在另一张桌子旁。这是我第一次见到他们。尤里看上去很壮实,而且很年轻。克拉拉举止优雅,但你能感觉到这是一个经历过好时光和坏时光的女人。对于一位年轻的寡妇来说,需要巨大的内在力量才能将加里培养成如今的巨大成功。加里是一个意志坚强的人,克拉拉在塑造他的性格方面可能发挥了很大的作用。克拉拉很友善,但也赢得了尊重。在加里的整个职业生涯中,每当有重大比赛情况时,她总是在那里提供支持。在费城的比赛中,克拉拉证明了对加里的镇定作用。

The morning after our arrival, while we were having breakfast in the hotel restaurant, Clara and Garry’s coach Yuri Dohokian were at another table. It was the first time that I had seen either of them. Yuri was stout looking and young. Clara was elegant in her demeanor, but you could sense that here was a woman who had seen both good times and bad. It took great inner strength for a young widow to raise Garry to the great success that he became. Garry is a strong-willed person, and Clara probably had a strong hand in shaping his personality. Clara was friendly, but she also commanded respect. Throughout Garry’s career, whenever there was a big match situation, she was always there to lend her support. For the Philadelphia match, Clara proved to have a calming influence on Garry.

赛前新闻发布会的出席人数还算不错,约有二十名记者到场。比赛颜色由深蓝在第一局比赛中以白色决定。在新闻发布会上,加里预测自己会获胜,但没有猜测最终比分。大多数人预测加里会轻松获胜;但事实并非如此。大卫·利维 (David Levy) 甚至预测 6-0 获胜。私下里,我认为大卫的结果是极不可能的,但我们也不是不可能输掉这场比赛。当时 IBM 以外的人都不知道我们的机器才使用了两周。我们还没有为比赛做好准备。

The press turnout at the pre-match press conference was decent, with around twenty reporters showing up. The playing color was decided with Deep Blue having White in the first game. At the press conference Garry predicted a win for himself but did not guess at the final score. Most were predicting an easy win for Garry; David Levy went so far as to predict a 6–0 wipeout win. Privately, I believed that David’s result was highly improbable, but it was not out of the question that we could lose the match very badly. No one outside IBM knew at the time that our machine was only two weeks old. We were nowhere near ready for the match.

首场比赛开始,深蓝队就充满了焦虑。我们自己从未见过深蓝玩过任何一款游戏,也不知道会发生什么。更糟糕的是,由于约克镇实验室的计算机网络问题,我们在启动游戏时遇到了一些早期故障。

The first match game began with a lot of anxiety for the Deep Blue team. We had never seen Deep Blue play a single game ourselves, and had no idea what to expect. Compounding this, we had some early glitches in starting the game owing to computer network problems in the Yorktown Lab.

在卡内基梅隆大学,在公共比赛期间,我们将 Deep Thought 工作站与部门本地网络隔离,以避免偶尔的网络中断,从而导致机器无法使用。最初,我们还计划将运行 Deep Blue 的 RS/6000 SP 与研究实验室本地网络的其余部分隔离。但该计划并不可行——这台机器对于整个实验室来说是半公开的,并且由于它的设置方式而无法被隔离。由于偶尔出现网络问题,深蓝在第一款游戏中需要一段时间才能启动。当我们还在等待深蓝启动时,比赛仲裁员 Mike Valvo 启动了深蓝的时钟。

At Carnegie Mellon, during public games we isolated the Deep Thought workstation from the departmental local network to avoid the occasional network outage, which could render the machine unusable. Originally, we also planned to isolate the RS/6000 SP running Deep Blue from the rest of the Research lab local network. But the plan was not workable—the machine was semi-public for the entire lab, and it could not be isolated because of the way that it was set up. As a result of the occasional network problems, Deep Blue took a little while to start up for the first game. Mike Valvo, the match arbiter, started Deep Blue’s clock while we were still waiting for Deep Blue to start up.

深蓝以 e4 开局,将棋子移至王面前两格。加里回应 c5,将后主教兵向前移动两个方格。这是西西里防御,加里最喜欢的对抗 e4 的防御。赛前,我们就他是否参加比赛进行了一些争论。乔尔认为加里在开场选秀中非常有纪律,他完全期待他扮演西西里人。在加里的c5之后​​,白棋有很多好的走法选择,但不幸的是,其中大多数都有大量的相关分析。乔尔只有几个月的时间来准备《深蓝》的开场白,所以我们选择播放较少人走的开场白。深蓝回答c3,将后主教兵向前移动一格。对于白方来说,这是一个合理但有些不规则的开局。优点是c3西西里很少有大师级别及7级以上的玩家玩,因此不需要做大量的开局准备。缺点是开局有点干,似乎没有为白方提供太多获胜机会。至少,那是我的意见基于我过去使用 Deep Thought 和 Deep Thought II 的经验,以及玩 c3 西西里岛双方的经验。乔尔有不同的看法。

Deep Blue opened the first game with e4, moving the pawn in front of the king by two squares. Garry responded with c5, moving the queen bishop pawn forward two squares. This was the Sicilian defense, Garry’s favorite defense against e4. Before the match, we had some debate about whether he would play it. Joel’s opinion was that Garry was very disciplined in opening selection and he fully expected him to play the Sicilian. There were many good move choices for White after Garry’s c5 but, unfortunately, most of them had reams of associated analysis. Joel only had months to prepare Deep Blue’s opening book, so we chose to play less traveled opening lines. Deep Blue answered c3, moving the queen bishop pawn forward one square. This is a sound, but somewhat irregular, opening for White. The advantage was that the c3 Sicilian was rarely played at the Grandmaster level and above 7, and as a result there was no need to do extensive opening preparation. The disadvantage was that the opening was a little bit dry, and did not seem to offer much in terms of winning chances for White. At least, that was my opinion based on my past experience with Deep Thought and Deep Thought II, playing both sides of the c3 Sicilian. Joel had a different opinion.

c3 西西里棋经常导致所谓的“孤立的后棋”位置,其中白方的后棋最终在相邻的文件上没有友方棋子。孤立的后兵通常在残局中是一个障碍,但在中局中它为白方提供了动态的进攻机会。问题在于,大多数国际象棋程序通常是良好的反动防御者,而不是良好的主动攻击者。乔尔花费了大量时间指导 Deep Blue Jr 如何从 c3 西西里岛的位置发起进攻。有一个棘手的问题。小深蓝不喜欢白棋的位置,并认为该位置对黑棋来说更好。通常,深蓝仍然会下正确的进攻棋,但有时它会允许黑方重复该位置,因为它希望黑方能够避免重复平局。所发生的情况是,1996年版的深蓝没有完全掌握攻击潜力的想法,因此做出了错误的评估。

The c3 Sicilian frequently leads to the so-called “isolated queen pawn” position, where White’s queen pawn ends up with no friendly pawn on the adjacent files. The isolated queen pawn is usually a handicap in endgames, but in middle games it offers dynamic attacking chances for White. The problem was that most chess programs were usually good reactionary defenders, but not good proactive attackers. Joel spent a significant amount of time coaching Deep Blue Jr how to attack from the positions arising out of the c3 Sicilian. There was one nagging problem. Deep Blue Jr did not like White’s positions, and assessed the position to be better for Black. Usually, Deep Blue would still play the right attacking moves, but occasionally it would allow Black to repeat the position since it expected Black to avoid the repetition draw. What happened was that the 1996 version of Deep Blue did not have full grasp of the idea of attacking potential, and hence made the wrong assessment.

事实上,深蓝没想到黑棋会重复这个位置,这在第一局比赛中创造了一个有趣的局面。在第十步中,加里施展了新的一步,将深蓝从开局书中剔除。三步之后,深蓝给了加里一个意想不到的选择。当我为深蓝下出第 13 步 Nb5 攻击他的皇后时,他大吃一惊。

The fact that Deep Blue did not expect Black to repeat the position created an intriguing situation in the first game. On move ten, Garry produced a new move that took Deep Blue out of its opening book. Three moves later, Deep Blue presented Garry with an unexpected choice. When I played move 13, Nb5 for Deep Blue attacking his queen, he was taken aback.

加里专注地看了一会儿棋盘,然后扬起眉毛偷偷看了我一眼。看到他的表情,我忍不住笑了。深蓝并没有真正“享受”它的地位,但加里却非常认真地对待这个地位。当我意识到自己在微笑时,我的笑容消失了。我不知道他如何解读我稍纵即逝的微笑,但他花了很长时间才回答。他的选择之一是通过将皇后移回 d8 来撤消他的最后一步 Qd6。在这种情况下,深蓝会通过下 Nc3 并重复该位置来撤销自己的最后一步,从而有效地提供平局。深蓝打出13.Nb5,部分原因是它愿意接受平局。加里没有将他的皇后移回d8,要么他不想提前抽牌,要么他没有意识到深蓝会提前抽牌。加里最终将他的皇后放在了e7上,深蓝认为这个方格有点尴尬,因为加里的一名骑士很快就被压在了他的皇后上。从此,他的地位不断走下坡路,走向绝望。

Garry looked at the board intently for a moment, then raised his eyebrows and stole a glance at me. I could not help but smile momentarily on seeing his expression. Deep Blue did not really “enjoy” its position, and yet Garry was taking the position very seriously. My smile faded the moment I realized that I was smiling. I don’t know how he interpreted my fleeting smile but he spent a long time on his reply. One of his choices was to undo his last move, Qd6, by moving his queen back to d8. In this case, Deep Blue would have undone its own last move by playing Nc3 and repeating the position, effectively offering a draw. Deep Blue played 13. Nb5, in part because it was willing to settle for a draw. Garry did not move his queen back to d8, either he did not want an early draw, or he did not realize that Deep Blue would have taken an early draw. Garry finally placed his queen on e7, a square that Deep Blue considered somewhat awkward, as one of Garry’s knights was soon pinned against his queen. From then on, his position steadily went downhill and became desperate.

位置在13.Nb5之后。

Position after 13. Nb5.

从那里开始,加里开始发动一场看起来令人讨厌的攻击,这肯定会让任何碳基国际象棋棋手感到不安。深蓝根本无视他的攻击,欢快地抓起后翼棋子。与此同时,随着时间的推移,加里的攻击看起来更加危险。他所需要的只是一脚自由走棋,比赛就会以对他有利的方式结束。就在这时,深蓝发动了自己的攻击。加里可以击退攻击,但会失去大部分材料并看到他的攻击消散。比赛的最后几步,他知道自己已经无望了,但距离比赛还有一个多小时的时间,深蓝还在花时间计算自己的最佳着法。深蓝第37步后,加里受够了,伸出了手,然后认输了。加里接下来的四步棋还剩大约五分钟,而深蓝接下来的三步棋还剩下大约一个小时。这是计算机首次在规定时间控制下战胜国际象棋世界冠军。历史被创造了!

From there, Garry proceeded to whip up a nasty looking attack that surely would have unnerved any carbon-based chess player. Deep Blue totally ignored his attack, and merrily snatched up queenside pawns. Garry’s attack, meanwhile, was looking more dangerous as time went on. All that he needed was one free move, and the game would be all over in his favor. At this point, Deep Blue unleashed its own attack. Garry could beat back the attack but would lose most of his material and see his attack dissipated. For the last few moves of the game, he knew that his position was hopeless, and yet Deep Blue, with over an hour of time left on its clock, was still taking time to calculate its best move. After Deep Blue’s move 37, Garry had had enough, offered his hand, and resigned. Garry had about five minutes left on his clock for his next four moves, while Deep Blue had about an hour left for its next three moves. It was the first win ever by a computer over a World Chess Champion under regulation time control. History was made!

与我握手后,Garry问的第一个问题是:“为什么花了这么长时间才出招?PC 程序会立即发现自己获胜。” 我耸耸肩,告诉他深蓝的时钟还剩太多时间了。接下来,他一脸痛苦地问道:“我哪里可以打得更好?” 这个问题让我大吃一惊。我几乎不会下国际象棋,而有史以来最伟大的国际象棋大师却问我他的棋艺可以在哪里改进。当然,加里的问题实际上是“深蓝认为我在哪里可以打得更好?” 我试图根据我对深蓝分析的记忆,尽可能地回答他的问题。加里感觉到我对这款游戏的看法与他和深蓝不同,几分钟后就放弃了,并迅速离开了大楼,没有与随行人员以外的任何人交谈。

After shaking hands with me, the first question that Garry asked was, “Why did it take so long to play its moves? A PC program would have seen that it was winning instantaneously.” I shrugged my shoulders and told him that Deep Blue simply had too much time left on its clock. Next, he asked with a pained look, “Where could I have played better?” This question took me by surprise. I could barely play chess and the greatest chess mind ever was asking me where he could have improved his play. Of course, Garry’s question really was “Where did Deep Blue think that I could have played better?” I tried to answer his question as best as I could from what I remembered of Deep Blue’s analysis. Sensing that I was not seeing the game on the same level as he and Deep Blue, Garry gave up after a few minutes, and quickly left the building without talking to anyone outside of his entourage.

直到乔尔·本杰明上台与我握手时,我才真正意识到这个场合有多么重要。乔尔笑着说,“我梦想着自己打败卡斯帕罗夫并与他握手,”并补充道,“这几乎一样好。” 整场比赛我都在控制自己的情绪。我现在感到既高兴又释然。默里和乔也欣喜若狂。一名记者过来索取报价,然后冲出去归档副本。当时周围的记者并不多,但接下来几天情况很快就发生了变化。

It did not really hit me how momentous the occasion was until Joel Benjamin came up to the stage to shake my hand. Joel said, with a broad smile, “I dreamt about beating Kasparov myself and shaking his hand,” adding, “This is almost as good.” Throughout the entire game I had been reining in my emotion. I now felt both elated and relieved. Murray and Joe were ecstatic as well. A reporter came by to get a quick quote and then rushed out to file the copy. There weren’t that many reporters around then, but that changed very quickly in the next few days.

第一场比赛期间,IBM 比赛网站就被淹没了。《今日美国》报纸在头版刊登了一篇文章,指出 IBM 的“深蓝”表现良好,但该网站却表现不佳。显然,网站管理员认为比赛期间只有几千名冲浪者会访问该网站。如果有人问我们,我们可以告诉网站管理员预计至少有数万名访问者,甚至可能有数十万8。乔纳森·谢弗 (Jonathan Schaeffer) 是网络访问者之一,他试图从加拿大埃德蒙顿的阿尔伯塔大学进行连接。由于无法接通,乔纳森浏览了新闻服务,发现路透社的一篇报道错误地称加里是白人。我们的朋友达普·哈特曼 (Dap Hartmann) 在波士顿,也遇到了连接问题。他和乔纳森同意定期尝试访问 IBM 站点并互相通报最新情况。他们能够通过不时获得棋盘位置来跟踪比赛。棋盘位置的网络显示没有说明哪个玩家有什么颜色,所以乔纳森和达普“看着”加里作为白色赢得了一场漂亮的胜利。正如预测的那样。故事结局。二十四小时后,乔纳森发现加里输了。对于第一天的许多网络访问者来说,乔纳森的情况可能非常典型。IBM 在第一场比赛中就开始动员网络团队。一台 IBM RS/6000 SP 超级计算机(与 Deep Blue 的主机类型相同)被选为新的 Web 服务器来处理工作负载。第 2 场比赛结束后,《今日美国》在头版刊登了另一篇文章,这次指出 IBM 已经修复了该网站。

The IBM match web site was swamped during the first game. The newspaper USA Today, carried a front page article stating that IBM’s Deep Blue was doing well but the web site was not. Apparently the web master had thought that only a few thousand surfers would be visiting the site during the game. Had we been asked, we could have told the web master to expect at least tens of thousands visitors and possibly hundreds of thousands 8. One of the web visitors was Jonathan Schaeffer, trying to connect from the University of Alberta in Edmonton, Canada. Unable to get through, Jonathan scanned the news service and found a Reuters story erroneously saying that Garry was White. Our friend Dap Hartmann, was in Boston and had problems connecting as well. He and Jonathan agreed to both try and access the IBM site periodically and keep each other posted. They were able to follow the game by getting the board position from time to time. The web display for the board position did not state which player had what color so Jonathan and Dap “watched” Garry score a nice win as White. As predicted. End of story. It was twenty-four hours later that Jonathan found out that Garry had lost. Jonathan’s situation was probably quite typical for a lot of web visitors on the first day. IBM started mobilizing the web team during the first game. An IBM RS/6000 SP supercomputer, the same type of machine as Deep Blue’s host computer, was enlisted as the new web server to handle the workload. After game 2, USA Today carried another front page article, this time stating that IBM had fixed the web site.

那天晚上,第一场比赛结束后,IBM 团队举行了庆祝晚宴。无论剩下的比赛发生了什么,我们都已经表明了自己的观点。卡斯帕罗夫阵营显然陷入了危机。据担任加里电脑顾问的弗雷德里克·弗里德尔 (Frederic Friedel) 说,加里在深夜在低于冰点的温度下在费城街道上散步。走着走着,他问道:“弗雷德里克,如果这东西无敌呢?”

That night after game one, the IBM team had a celebration dinner. No matter what happened in the remainder of the match, we had made our point. The Kasparov camp apparently was in a crisis. According to Frederic Friedel, who was serving as Garry’s computer consultant, Garry went for a late night walk on the Philadelphia streets in below freezing temperatures. During the walk, he asked, “Frederic, what if this thing is invincible?”

质量源于数量

Quality from Quantity

乔尔和穆雷在晚上和第二天早上一起练习了深蓝第二场比赛的开局动作。我参加了开幕式准备会议的一部分。与开盘相关的测试头寸被手工输入到测试文件中,深蓝连夜分析测试文件中的头寸,以便第二天早上进行人工分析。我们都同意,如果要再举行一场比赛,我们需要自动化开场准备过程,既可以加快速度,又可以减少人为错误的机会。我们不知道我们的观察有多么有预见性。开场准备会议结束后,Murray 将在他的 IBM Thinkpad 笔记本电脑上准备的开场文件上传到我们约克镇实验室的深蓝。或者说他是这么想的。

Joel and Murray worked together on Deep Blue’s opening moves for game two during the night and the next morning. I was present during part of the opening preparation session. Test positions related to the opening were entered into a test file by hand, and Deep Blue analyzed the positions in the test file overnight for human analysis the next morning. We all agreed that if there was ever going to be another match, we would need to automate the opening preparation process, both to speed it up, and to reduce the chance for human error. We had no idea how prophetic our observation was. After the opening preparation session, Murray uploaded the opening file, which was prepared on his IBM Thinkpad laptop computer, to Deep Blue in our Yorktown lab. Or so he thought.

第二场比赛开始得很天真。第二步,深蓝就出局了。我当时并不知道,但在上传开局文件时出了问题,深蓝没有新的开局书。由于我们从未使用过新的开局书,为第二场比赛准备的开场最终在第四场比赛中进行。然而,《深蓝》只是表面上出现在本书中。从《深度思考II》开始,默里就创作了我们所说的“扩展书”,试图捕捉人类“开放理论”的概念。当深蓝退出传统意义上的开局后,它开始使用大师对局的统计数据来帮助它决定走哪一步。如果某个位置的某一招式是由强大的宗师下出并取得了很好的成绩,那么该招式将按比例获得奖励结果有多好。这使得深蓝偏向于玩“已知的好路线”,但仍然让它有机会玩出意想不到的棋步,在它的评估中,这些棋步似乎提供了更好的机会。

Game two started innocently enough. On the second move, Deep Blue was out of its opening book. I did not know it at the time, but something had gone wrong when the opening file was uploaded, and Deep Blue did not have the new opening book. Since we never did use the new opening book, the opening prepared for game two ended up being played in game four. Deep Blue was only out of the opening book superficially, however. Ever since Deep Thought II, Murray had created what we called the “extended book”, which tried to capture the human concept of “opening theory”. When Deep Blue was out of the opening book in the traditional sense, it started to use statistics from games played by Grandmasters to help it decide which move to play. If a move in a particular position had been played by strong Grandmasters with great results, then the move would be given a bonus, proportional to how good the results were. This gave Deep Blue a bias to play the “known good lines”, but still allowed it a chance to play unexpected moves that, in its assessment, appeared to give better chances.

穆雷在传输文件时出现的人为错误并不是本场比赛中的第一次。第六步,深蓝有可能吃掉两次棋子。我在官方棋盘上错误地走错了棋子。默里冲进去向瓦尔沃、仲裁员和我通报了这个错误。很快它就得到了纠正,时间又回到了加里的时钟上。扩展书在这个游戏中效果很好。当深蓝留下已知的开局理论时,加里仅拥有微弱的优势。具有讽刺意味的是,穆雷的人为错误与他早期在扩展书上的工作相结合,为深蓝创造了一个非常可玩的位置。

Murray’s human error in transferring the file was not the first in this game. On move six, two pawn captures were possible for Deep Blue. I mistakenly played the wrong pawn capture on the official chessboard. Murray rushed in to inform Valvo, the arbiter, and me about the mistake. It was soon corrected, and time was added back to Garry’s clock. The extended book worked quite well in this game. By the time that Deep Blue left known opening theory, Garry had only a slight advantage. Murray’s human error, ironically, in combination with his earlier work on the extended book, produced a very playable position for Deep Blue.

第十八步,加里提出了假棋子牺牲,深蓝拒绝了。拿走棋子会给他带来强烈的攻击。下一步,他祭出了真正的典当。赛后分析表明,如果深蓝下了最好的棋,加里的棋子牺牲应该会导致平局。比赛结束后,我们在分析问题所在时,发现了深蓝评测软件的一个“bug”。深蓝国际象棋芯片包含一个硬件,可以识别所谓“前哨”方格上的小棋子(主教和马),这些方格上的小棋子受到友方棋子的保护,并且不再受到对手棋子的攻击。棋片进一步确定小棋子是否不能再被对手的小棋子挑战(或交换)。在国际象棋传说中众所周知,如果一个前哨的马不再受到挑战,那么它就可以像车一样有价值。然而,无可争议的外派主教或许并不比普通的外派主教好到哪里去。不幸的是,分配这些值的软件代码错误地为不可挑战的主教分配了与不可挑战的外派骑士相同的值。软件评估中的这个严重错误困扰着深蓝在第二场比赛的剩余时间里的表现。如果我们有更多的时间准备,就像我们在深蓝复赛中所做的那样,这个相当明显的“bug”早就在测试游戏中被发现了。

On move eighteen, Garry offered a sham pawn sacrifice and Deep Blue declined. Taking the pawn would have given him a strong attack. On the next move, he offered a real pawn sacrifice. Post-game analysis indicated that Garry’s pawn sacrifice should have led to a draw if Deep Blue had played the best moves. After the game, while trying to analyze what went wrong, we found an evaluation software “bug” in Deep Blue. The Deep Blue chess chip contained a piece of hardware that recognized minor pieces (bishops and knights) on so-called “outpost” squares, squares on which they were protected by friendly pawn(s) and could no longer be attacked by the opponent’s pawns. The chess chip further identified whether the minor pieces could no longer be challenged by (or exchanged with) the opponent’s minor piece. It is well known in chess lore that if an outposted knight can no longer be challenged, then it can be as valuable as a rook. However, unchallengeable outposted bishops are perhaps no better than regular outposted bishops. Unfortunately, the software code that assigned the values, wrongly assigned unchallengeable bishops the same values as unchallengeable outposted knights. This gross error in the software evaluation plagued Deep Blue’s play for the remainder of game two. If we had had more time for preparation, as we did in the Deep Blue rematch, this fairly obvious “bug” would have been discovered during test games long beforehand.

加里牺牲棋子后,发生了一系列的棋子交换。在这轮比赛中,深蓝至少有两次相对轻松的平局机会。由于渴望,它错过了第一个帕克是其“不可挑战的”外派主教。第二次失败是因为该计划存在避免成为交易女王的偏见。进入1996年的比赛,我们认为Garry绝对是更强的选手,为了增加我们的战术机会,我们对深蓝交易后的行为给予了很大的惩罚。皇后贸易回避偏见在这场游戏中适得其反。加里在第一手观察到这种偏见后,在后来的比赛中巧妙地利用了它。

A sequence of piece exchanges ensued after Garry’s pawn sacrifice. Deep Blue had at least two chances for relatively easy draws during the sequence. The first one it missed because of its desire to park its “unchallengeable” outposted bishop. The second one it missed because of a bias placed into the program to avoid trading queens. Going into the 1996 match, we believed that Garry was definitely the stronger player, and to increase our tactical chances, we gave Deep Blue a large penalty for trading the queens. The queen trade avoidance bias backfired in this game. Garry, after observing the bias first hand, exploited it masterfully in later games.

尘埃落定后,加里明显占据优势,但深蓝仍有平局机会。据深蓝称,他随后允许将自己最危险的棋子换成深蓝较弱的棋子之一,这是不必要的。棋盘上出现了一种结局,双方各有一个皇后和一个主教,主教位于相反颜色的方块上。结局大概是平局了,但是深蓝的位置很难守住9。加里发起了一场漫长而微妙的行动,最终为他赢得了一枚棋子。随着典当赤字,深蓝的地位开始崩溃。在第73手,第一次控制后接替我担任操作员的穆雷向深蓝辞职。

After the dust settled, Garry had a clear edge, but Deep Blue still had drawing chances. He then allowed his most dangerous pawn to be traded for one of Deep Blue’s weak pawns, unnecessarily, according to Deep Blue. An ending with each side having a queen and a bishop, with the bishops on opposite colored squares, appeared on the board. The ending was probably drawn, but Deep Blue’s position was very difficult to hold 9. Garry initiated a long and subtle maneuver that netted him a pawn. With the pawn deficit, Deep Blue’s position began to crumble. On move 73, Murray, who had replaced me as the operator after the first time control, resigned for Deep Blue.

加里大喜,立即前往解说室,对观众讲了三十多分钟。克拉拉示意他打住,但在兴奋的状态下,他对比赛进行了详细的事后分析,并对迄今为止的比赛进行了总结。观众们非常享受。他在总结中说道:“我祝贺 IBM 的研究人员取得了巨大的成就。他们成功地将数量转化为质量。” 这有助于减轻我们刚刚遭受的损失所带来的痛苦。接下来的赞美让我们大吃一惊。“在某些位置上,它的洞察力如此之深,以至于它的表现就像上帝一样。”

Garry was exultant, went immediately to the commentary room, and spoke to the audience for over thirty minutes. Clara signaled him to cut it short, but in his excited state he went into a detailed post mortem analysis of the game, and gave a summary of the match so far. The audience thoroughly enjoyed it. During the summary, he said, “I congratulate the researchers at IBM for a fantastic achievement. They have succeeded in converting quantity into quality.” That helped reduce the sting from the loss that we just suffered. The next praise took us by surprise. “In certain kinds of position, it sees so deeply that it plays like God.”

那天晚上,我们仔细地检查了游戏,试图找出问题所在。当穆雷发现外派主教的问题时,我们都感到震惊。令人惊讶的并不是我们有一个严重的错误——考虑到我们准备比赛的时间很短,我们预计会有更多的错误。最大的惊喜是我们距离赢得比赛有多近。如果没有这个错误,深蓝很可能会平局第二场。在第一场比赛之后,加里就已经产生了很多自我怀疑。A第二场平局可能会让比赛赔率对我们有利,不仅因为我们会领先一场,还因为加里会承受心理压力。

That night, we went over the game carefully trying to find out what went wrong. When Murray discovered the problem with the outposted bishop, we were all shocked. The surprise was not from the fact that we had a serious bug—we expected many more bugs, given the little time that we had to prepare for the match. The big surprise was how close we had been to possibly winning the match. Without the bug, Deep Blue might very well have drawn game two. Garry already had lots of self-doubt after game one. A drawn game two could have tilted the match odds heavily in our favor, not just because we would remain up by one game, but also because of the psychological pressure that would be placed on Garry.

第二天,2 月 12 日,理论上是“休息”日,尽管我们,可能还有他们,都在努力工作。加里阵营的两名成员确实得到了休息。特里·菲尼克斯 (Terrie Phoenix) 带着克拉拉·卡斯帕洛娃 (Clara Kasparova) 和弗雷德里克·弗里德尔 (Frederic Friedel) 去购物。1996年的比赛中,两队保持着友好的关系,特里可能在缓和局势方面发挥了主要作用。特里在 1997 年复赛前离开了 ACM。1997 年的重赛中,两队之间的关系变得更加激烈,部分原因是深蓝很有可能赢得比赛,但像特里这样的人缺席也很可能也是原因之一。

The next day, February 12, was in theory a “rest” day although we, and probably they, were hard at work. Two members of Garry’s camp did get some rest. Terrie Phoenix took Clara Kasparova and Frederic Friedel shopping. During the 1996 match, the two teams maintained a cordial relationship, and Terrie probably played the major role in smoothing things along. Terrie left ACM before the 1997 rematch. The relationship between the teams was more competitive at the 1997 rematch, in part because of the very real possibility of Deep Blue winning the match, but the absence of someone like Terrie most likely played a part as well.

深蓝团队很快就安排好了这个休息日的日程。穆雷和乔尔的工作压力最大,为接下来的两场比赛做准备。我花时间进行测试,有时会发现国际象棋程序中的问题。程序本身足够稳定,大部分问题都与残局有关。乔的行为超出了消防员的职责范围,他在出现问题时解决了问题。休息日他还有另一项任务——加快程序进度。IBM RS/6000 SP 上的并行国际象棋程序比在单个 RS/6000 工作站上运行的 Deep Blue Jr 复杂得多。这也是一个相当新的程序,还有很大的改进空间。一些进步是在整个比赛期间完成的,但大部分是在休息日期间实现的。

The Deep Blue team quickly settled into a routine for this rest day. Murray and Joel had the most stressful work, preparing the opening for the next two games. I spent my time running tests, which sometimes revealed problems in the chess program. The program itself was sufficiently stable, and most of the problems were endgame related. Joe acted beyond the call of duty as the fire fighter who fixed up the problems as they came along. He had another task during rest days—speeding up the program. The parallel chess program on the IBM RS/6000 SP was quite a bit more complicated than Deep Blue Jr, which ran on a single RS/6000 workstation. It was also a fairly new program, and there was lots of room for improvement. Some of the improvement was done throughout the match, but most of it came during the rest days.

偶尔,我会把自己从测试中抽离出来,加入穆雷和乔尔的行列,讨论我们的开局选择。深蓝赢得了第一局,但乔尔和穆雷都对开局的结果感到不满意;加里本可以轻松重复抽签。在一场短暂的比赛中,每次有白棋时,你都必须努力获胜。我们需要更好的东西。穆雷和乔尔为了开局问题一直工作到深夜。乔也因为自己的任务而熬夜。我决定休息一下,并与 IBM 的其他人员一起享用了一顿丰盛的晚餐。我喜欢吃,任何工作都不能阻止我享受一顿美餐。

Once in a while, I would remove myself from the testing and join Murray and Joel in their discussions about our opening choices. Deep Blue won the first game, but neither Joel nor Murray was happy with how the opening went; Garry could have had an easy repetition draw. In a short match, you have to try to win every time that you have White. We needed something better. Murray and Joel worked late into the night on the opening problems. Joe also stayed late on his own task. I decided to take a break and joined the rest of the people from IBM for a nice dinner. I like to eat, and no work was going to stop me from enjoying a good meal.

穆雷和乔尔在休息日并没有想出一个很好的办法来解决我们的开局问题。但在比赛第三场比赛的前一天早上,乔尔宣称他在晚上得到了灵感。深蓝会提前偏离其在第一场比赛中所采取的动作。加里可以演奏几种可能的延续。我们检查了一些,它们看起来很适合深蓝。在其中一些案例中,深蓝找到了明显的获胜路线。这看起来棒极了!没有足够的时间检查所有内容,但我们已尽我们所能并添加了新的开场白。

Murray and Joel did not come up with a good solution to our opening problems during the rest day. But the morning before the third game of the match, Joel declared that he had had an inspiration during the night. Deep Blue would deviate early from the moves it played in game one. There were several possible continuations that could be played by Garry. We checked out a few, and they looked good for Deep Blue. In some of them, Deep Blue found clearly winning lines for itself. This was looking great! There was not enough time to check everything out but we did as much as we could and put in the new opening lines.

一切都按预期进行;加里(Garry)发挥了我们准备好的开场白。第十二步之后,深蓝就脱离了我们检查的书籍行,但仍然处于我们预期的序列之一中。第十八手,乔尔预计深蓝会下 18.Be5,并建立了潜在的进攻。深蓝拒绝使用 18.Be5,而是移动了其中一辆车。原来,Garry对18.Be5进行了强有力的反驳,他在棋盘后面的几步中就看到了反驳。没有18.Be5,深蓝没有攻击,只有几个弱棋子必须立即防守,或者至少看起来是这样。虽然深蓝的位置看起来很严峻,但这只是幻觉。经过几个非人的举动后,很明显深蓝并没有输,甚至可能还稍稍领先。第39手,双方都认为继续比赛没有意义,并商定平局。加里和深蓝的这场比赛都令人印象深刻,但我们浪费了一个白棋。

Everything went as expected; Garry played into our prepared opening lines. After move twelve, Deep Blue was out of our checked book lines, but was still in one of our expected sequences. On move eighteen, Joel expected Deep Blue to play 18. Be5, and set up a potential attack. Deep Blue refused to play 18. Be5, and moved one of its rooks instead. It turned out that Garry had a strong refutation against 18. Be5, and he had seen the refutation several moves back at the board. Without 18. Be5, Deep Blue had no attack but several weak pawns that had to be defended immediately, or at least that was the way it looked. While Deep Blue’s position looked optically grim, it was an illusion. After several inhuman looking moves, it became clear that Deep Blue was not losing, and might even be slightly ahead. On move 39, both sides agreed that there was no point in playing on, and a draw was agreed. It was an impressive game by both Garry and Deep Blue, but we had wasted one of our Whites.

由于第二场比赛的开场准备工作从未进行过,因此第四场比赛之前的早上没有进行开场准备工作。这一次,深蓝打出了斯拉夫防守。深蓝在离开第一本书后的位置结果不如它在第二场比赛中独自达到的位置。在第十五步中,加里有机会做出投机性的牺牲,但他决定放弃。赛后,Garry拜访了IBM研究院,与深蓝一起分析了形势。深蓝认为这种牺牲是合理的,但他并没有明显的胜利。看到深蓝想出的一些进攻和防守招式后,加里很庆幸自己没有做出牺牲。然而,他说他仍然会与任何其他对手(无论是人类还是电子对手)进行牺牲。考虑到加里在第十五手的位置,深蓝本身就会做出牺牲。

Since the opening prepared for game two was never played, there was no opening preparation session in the morning before game four. This time, Deep Blue played a Slav defense. Deep Blue’s position after leaving its opening book turned out to be less favorable than the one it reached on its own in game two. On move fifteen, Garry had a chance to make a speculative sacrifice, but decided against it. After the match, Garry visited IBM Research and analyzed the position with Deep Blue. Deep Blue assessed the sacrifice as sound, but there was no clear win for him. After seeing some of the attacking and defensive moves that Deep Blue came up with, Garry was glad that he did not make the sacrifice. However, he said that he would still have played the sacrifice against any other opponent, human or electronic. Deep Blue would have played the sacrifice itself, given Garry’s position at move fifteen.

深蓝下完第21步后,我离开了牌桌,冲向休息室。比赛规则规定,只有当深蓝在移动或刚刚移动时,操作员才能离开或返回桌子。这有点不寻常。在人对人的游戏,任何一个玩家都可以随时离开桌子。合同谈判时我们并没有认为这个让步太大,就同意了这个条款。不管怎样,当我冲回游戏室时,加里还在思考,我不被允许回到牌桌。他想了很长时间他的二十秒动作,我等着他做出动作。由于接下来发生的事情,我们试图改变 1997 年复赛的规则,允许操作员安静地进出。最初规则发生了变化,但不知何故最终还是采用了旧规则。

After making Deep Blue’s move 21, I left the playing table and rushed to the rest room. The match rules specified that the operators could leave or return to the table only when Deep Blue was on move, or it had just moved. This was somewhat unusual. In human vs. human games, either player can leave the table at any time. We did not consider the concession too great when the contract was negotiated, and had agreed to the term. Anyway, by the time I rushed back to the game room, Garry was still thinking and I was not allowed back to the table. He thought for a long time on his twenty-second move, and I waited until he made the move. As a result of what happened next, we tried to change the rules for the 1997 rematch to allow operators to come and go quietly. The rules were changed initially, but somehow the old rules were adopted in the end.

加里·卡斯帕罗夫 (Garry Kasparov) 下棋 22. f5 后的位置。

Position after Garry Kasparov’s move 22. f5.

当我回到桌子旁时,我的电脑屏幕已进入空白的“屏幕保护”模式。我输入了加里的招式 f5,深蓝崩溃了。屏幕保护模式将 f5 中的“f”解释为唤醒字符,深蓝收到“5”而不是完整的移动 f5。Joe 一直在开发深蓝并行国际象棋程序,并在程序中添加了一些测试命令来帮助他完成工作。不幸的是,“5”是测试命令之一,告诉 5 号工作站监听下一个命令。不用说,如果不在程序测试模式下执行命令“5”,则会导致程序立即崩溃。

When I got back to the table, my computer screen had entered into a blank “screen saver” mode. I typed in Garry’s move f5, and Deep Blue crashed. The screen saver mode interpreted the ‘f’ in f5 as a wakeup character, and Deep Blue received a ‘5’ instead of the full move f5. Joe had been working on the Deep Blue parallel chess program, and had added some testing commands to the program to help him with his work. Unfortunately, ‘5’ was one of the testing commands telling workstation number 5 to listen to the next command. Needless to say, the command ‘5’ would cause an immediate program crash when not executed in the program-testing mode.

加里正在去休息的路上。程序崩溃使我必须在时钟滴答作响的情况下重新启动深蓝离开。我通知了仲裁员,并开始尽可能快地打字。穆雷和乔冲进游戏室查看出了什么问题。加里看到了这些活动,很慌张,并向仲裁员投诉。克拉拉整场比赛都在游戏室里,她用俄语对加里喊了几句,他平静下来,也许意识到深蓝的时钟已经滴答作响。(事后有人告诉我,克拉拉大喊“闭嘴”。) 坠机回来后,深蓝很快就回答了Bxc4。加里原本预计会下Bxf5,之后他准备下Nxf7,牺牲他的马换取两个棋子。一份报告声称,加里声称骑士的牺牲将使白棋在深蓝的搜索范围之外获得胜利。从国际象棋书籍中对这场比赛的错误分析来看,骑士的牺牲肯定会战胜人类的对手。深蓝的赛后分析表明,黑方的防守走法比书中给出的走法更强,而国际象棋书中给出的一些进攻走法似乎让白方彻底失败。就我个人而言,我相信这种牺牲会导致双方打出最好的平局,因为有足够的制衡因素。加里赛后访问IBM研究院期间,他也想看看深蓝在f5之后的分析。1996 年的深蓝更喜欢 Bxc4,略高于 Bxf5。1997年版本的深蓝首选Bxf5。如果像我怀疑的那样,Nxf7 在 Bxf5 之后不起作用,那么 Bxc4 和 Bxf5 都可以玩。

Garry was on his way to take a break. The program crash made it necessary for me to restart Deep Blue, with our clock ticking away. I informed the arbiter, and started typing as fast as I could. Murray and Joe rushed into the game room to find out what went wrong. Garry saw the activities, got flustered, and complained to the arbiter. Clara, who was in the game room throughout the entire match, shouted something in Russian to Garry, and he calmed down, perhaps realizing that Deep Blue’s clock had been ticking. (Someone told me afterwards that Clara had shouted “shut up”.) After coming back from the crash, Deep Blue quickly answered Bxc4. Garry had expected Bxf5, after which he had prepared to play Nxf7, sacrificing his knight for two pawns. One report asserted that Garry claimed the knight sacrifice would have led to a winning position for White beyond Deep Blue’s search horizon. Judging from the faulty analyses given in chess books on the match, the knight sacrifice would certainly have won against human opposition. Post-game analyses with Deep Blue suggested stronger defensive moves for Black than the ones given in the books, and some of the attacking moves given in chess books appear to lose outright for White. Personally, I believe the sacrifice would have led to a draw with best play by both sides as there were enough counterbalancing factors. During Garry’s visit to IBM Research after the match, he also wanted to see Deep Blue’s analysis after his f5. The 1996 Deep Blue preferred Bxc4 slightly over Bxf5. The 1997 version of Deep Blue preferred Bxf5. If Nxf7 does not work after Bxf5 as I suspect, then both Bxc4 and Bxf5 are playable.

当深蓝打出Bxc4来回应f5时,Garry的开局优势大部分已经消失。即便如此,深蓝的地位仍然显得脆弱。加里第 34 步棋后,深蓝将其弱 c 棋子向前推进一格,突然间,它对将死者产生了威胁。在第40步时间控制之前,加里每步棋还剩下不到一分钟的时间。接下来的几步,双方都至少出现了一处失误。加里之所以这么做是因为他的游戏时间很短,而深蓝的则是因为我们的程序开发时间很短。比赛时,国际象棋程序仅使用了国际象棋芯片识别的位置特征的三分之一。该游戏的一个关键的残局特征是,如果所有棋子都在棋盘的同一半上,则游戏更有可能平局。国际象棋芯片可以识别这一重要的残局特征,但软件没有利用这种能力。深蓝出手让加里移除了最后一个后侧棋子。深蓝还是占据了更好的位置,但他毫不费力地找到了最好的出路,并为深蓝的强大骑士牺牲了他的车。第 50 步,双方同意平局。四场比赛过后,双方战成两分。

By the time Deep Blue played Bxc4 in response to f5, most of Garry’s opening advantage had dissipated. Even so, Deep Blue’s position looked tenuous. After Garry’s move 34, Deep Blue pushed its weak c pawn forward one square and suddenly it was threatening checkmate. Garry had less than one minute per move left before the time control at move 40. For the next few moves, both sides made at least one mistake. Garry made his because he was short on game time, and Deep Blue’s because we were short on program development time. At the time of the match, only one third of the positional features recognized by the chess chip were in use by the chess program. One key endgame feature for this game was, that if all the pawns are on the same half of the chessboard, the game is more likely to be drawn. The chess chip could recognize this important endgame feature, but the software did not make use of this ability. Deep Blue made a move that allowed Garry to remove its last queenside pawn. Deep Blue still had the better position, but he had no problem finding the best way out and sacrificed his rook for Deep Blue’s strong knight. On move 50, a draw was agreed. After four games, the match was tied at two points each.

在接下来的休息日,我在酒店电梯里看到了弗雷德里克·弗里德尔。他描述了加里的遭遇,“他回到酒店房间,脱光了内裤,盯着天花板看了很长时间。”

During the next rest day, I saw Frederic Friedel in the hotel elevator. He described what happened to Garry, “He went back to the hotel room, stripped to his underpants, and stared at the ceiling for a long time.”

然而,加里很快就会康复。

Garry, however, would soon recover.

不断增长的珠穆朗玛峰

A Mount Everest That Grows

创造一台能够在比赛中击败世界象棋冠军的计算机是计算机科学中最古老的圣杯之一。这一探索可以追溯到 18 世纪,早于 19 世纪查尔斯·巴贝奇 (Charles Babbage) 的机械计算机,也早于第一台可编程电子计算机 ENIAC。1949 年,即 ENIAC 创建三年后,信息论之父克劳德·香农 (Claude Shannon) 重新点燃了探索之火,并描述了如何让计算机下棋。在克劳德做出开创性的工作之后,这个圣杯被隐藏了近 50 年。对于计算机科学家来说,击败国际象棋世界冠军的追求对于登山者来说就像攀登珠穆朗玛峰一样重要。

The quest to create a computer that could beat the World Chess Champion in a match, was one of the oldest holy grails in computer science. The quest could be traced back to the 18th century, before Charles Babbage’s mechanical computers in the 19th century, and well before ENIAC, the first programmable electronic computer. In 1949, three years after the creation of ENIAC, Claude Shannon, the father of information theory, re-ignited the fire for the quest and described how a computer could be made to play chess. The holy grail remained hidden for nearly 50 years after Claude’s seminal work. To a computer scientist, the quest to beat the World Chess Champion had taken on the same significance as climbing Mount Everest to a mountain climber.

有一点细微的差别。国际象棋世界冠军是一个活人。我们正在谈论一座正在生长的珠穆朗玛峰。

One slight difference. The World Chess Champion is a living being. We are talking about a Mount Everest that grows.

1996 年比赛的前四场比赛让计算机科学家一睹了圣杯的风采。我们可以看到珠穆朗玛峰的顶峰;问题是,这座珠穆朗玛峰很快就会高出几百英尺,就在世人的眼前。

The first four games of the 1996 match gave computer scientists a glimpse of the holy grail. We could see the top of Mount Everest; the problem was, that this Mount Everest would soon grow a few hundred feet taller, right in front of the eyes of the world.

2 月 15 日,比赛的最后一个休息日,乔尔和穆雷继续为怀特寻找可以为我们带来另一场胜利的开场线。到了下午,我们找到了一条似乎可以为深蓝对抗加里的西西里防守提供良好机会的防线。仍然对第三场比赛的失败感到痛苦,这一次我们在休息日开始检查新线路。第二天早上,我们回顾了深蓝对可能的延续的分析,发现并修补了一些漏洞。

On 15 February, the last rest day of the match, Joel and Murray continued their search for an opening line for White that could give us another win. By the afternoon, we had found a line that appeared to give Deep Blue good chances against Garry’s Sicilian defense. Still smarting from the debacle in game three, this time we started checking the new line during the rest day. The next morning, going over Deep Blue’s analyses of the possible continuations, we found and patched a few holes.

深蓝采取了与第一局和第三局相同的开局棋,e4。Garry立即回复e5。我再三考虑。没有西西里岛!我们让他放弃了他的西西里岛防御!他害怕我们对待 c3 西西里岛人的方式吗?冷静下来。这也意味着我们最后一天半的开场准备就白费了。两支球队现在都处于新的领域。或者以深蓝为例,两个月前的领土10。比赛前两个月,穆雷花了几个小时在书中添加了一些替代台词,以防加里演奏其他内容。此后,对这些替代产品线进行了很少的测试。

Deep Blue played the same opening move as it did in game one and game three, e4. Garry instantly replied e5. I did a double take. No Sicilian! We had gotten him to abandon his Sicilian defense! Was he afraid of our treatment of the c3 Sicilian? Calm down. This also meant that our opening preparations for the last day and half were wasted. Both teams were in new territory now. Or in the case of Deep Blue, two-month old territory10. Two months before the match, Murray spent a few hours putting in some alternative book lines, just in case that Garry played something else. Very little testing had been done with these alternative lines since.

接下来的比赛是Nf3和Nf6。加里想打佩特罗夫防守,也称为俄罗斯开局。众所周知,彼得罗夫有点爱画画。深蓝避开了佩特罗夫,进入了四骑士苏格兰威士忌的开局。事实上,我们从未测试过开局,这一事实从深蓝玩游戏的方式就可以看出。到第23步时,加里取得了微弱的优势;成功后,他提出了平局。

The game followed with Nf3 and Nf6. Garry wanted to play the Petroff defense, also known as the Russian opening. Petroff is known to be somewhat drawish. Deep Blue avoided the Petroff and went into the Four Knights Scotch opening. The fact that we had never tested the opening showed in the way Deep Blue played the game. By move 23, Garry had gained a slight advantage; after making it, he offered a draw.

加里·卡斯帕罗夫 (Garry Kasparov) 下棋后的位置为 23。

Position after Garry Kasparov’s move 23.

我是董事会的操作员。深蓝队在赛前就已经同意,将平局报价转达给后台的其他队员,然后再做出决定。从那里11 . 深蓝团队中没有人预料到会在这么早的阶段出现平局。之前的两场抽签一直进行到分出明确的抽签位置为止。我给后面的房间打电话,把这个提议告诉了大家。

I was the operator at the board. The Deep Blue team had agreed before the match that a draw offer would be relayed to the rest of the team in the back room, and the decision would be made from there11. No one on the Deep Blue team was expecting a draw to be offered at this early stage. The two earlier draws had been played until clearly drawn positions were reached. I rang the back room and told the gang about the offer.

深蓝评估认为加里的处境稍好一些。我对董事会的职位有点紧张。加里在王翼上多了一颗棋子,这挤压了深蓝的位置。早在《Deep Thought II》的时候,我就曾经玩过一场游戏,其中一方拥有如此多的王翼棋子,并利用多余的棋子发起了致命的攻击,效果非常好。Deep Thought II 在理解王侧兵多数在攻击国王时的价值方面存在严重问题。深蓝有可能理解棋盘上的位置,但我怀疑软件编码还没有完成。加里在旁边看着,我无法向团队表达我的担忧。团队的其他成员必须做出决定。此外,在比赛还剩下很多比赛的情况下,这么早就接受平局似乎是错误的做法。你并不是每天都能在一场严肃的比赛中与国际象棋世界冠军坐在一起。里屋没有做出任何决定。与此同时,深蓝已经决定了自己想玩什么,但并不知道抽签提议。仲裁者迈克·瓦尔沃 (Mike Valvo) 迅速做出裁决,坚持认为我们要么必须立即接受抽签提议,要么采取深蓝的预期行动。迈克的突然裁决令人震惊,即使在今天我仍然认为它有些值得怀疑。在人类玩家之间的国际象棋游戏中,平局提议一直有效,直到被拒绝或在棋盘上采取行动为止。棋盘上没有任何动作,比赛规则明确规定,无论是否咨询深蓝,操作者都可以接受平局。换句话说,当抽签提议提出时,决定权就不再是深蓝的了,深蓝已经决定继续比赛这一事实不应影响运营商如何或何时做出决定。迈克·瓦尔沃的裁决迫使我们立即做出决定,抽签提议被拒绝。

Deep Blue assessed the position as slightly better for Garry. I was a little nervous about the position on the board. Garry had one extra pawn on the king wing, and it was cramping Deep Blue’s position. Back in the Deep Thought II days, I once played through a game where one side had such a kingside pawn majority, and launched a deadly attack using the extra pawn to great effect. Deep Thought II had serious problems understanding the value of the kingside pawn majority in attacks against the king. Deep Blue had the potential to understand the position on the chessboard, but I suspected that the software coding had not been done yet. With Garry looking on, I could not voice my concern to the team. The rest of the team would have to make the decision. Besides, accepting a draw this early, when there was still a lot of play left in the game, seemed like the wrong thing to do. It was not every day that you get to sit across from the World Chess Champion in a serious game. No decision came from the back room. Meanwhile, Deep Blue had made up its mind on what it wanted to play, unaware of the draw offer. Mike Valvo, the arbiter, made a snap ruling insisting that we either had to accept the draw offer immediately or make Deep Blue’s intended move. Mike’s sudden ruling came as a shock, and even today I consider it somewhat questionable. In chess games between human players, the draw offer stands until it is declined or until a move is made on the chessboard. No move had been made on the chessboard, and the match rule explicitly stated that the operator could accept the draw offer with or without consulting Deep Blue. In other words, when the draw offer was made, the decision was no longer Deep Blue’s, and the fact that Deep Blue had made up its mind on the move to play, should not have affected how or when the operator made the decision. Mike Valvo’s ruling forced us to make up our mind immediately, and the draw offer was declined.

接下来的事情并不美好,除非你是加里。随着比赛的展开,深蓝显示出一些弱点。一些弱点可以通过简单地激活国际象棋芯片上未使用的硬件评估项来修复。其他人则需要在硬件或软件中添加新的评估术语。在第 47 手,也就是卡斯帕罗夫提出和牌后的二十四手,我们向深蓝认输。辞职时,克拉拉·卡斯帕罗娃明显松了一口气。加里只需要在最后一场比赛中打平就可以赢得比赛,并且他将拥有怀特。

What then followed was not pretty, unless you were Garry. Deep Blue showed several weaknesses as the game unfolded. Some of the weaknesses could be fixed by simply activating the unused hardware evaluation terms on the chess chip. Others required adding new evaluation terms either in hardware or in software. On move 47, twenty-four moves after Kasparov’s draw offer, we resigned for Deep Blue. At the time the resignation took place, Clara Kasparova gave a visible sigh of relief. Garry just needed one draw for the last game to win the match, and he would have White.

那天晚上和第二天早上,默里和乔尔努力工作。为了打平,我们必须以黑方赢得最后一局。这是一个艰巨的任务。加里在他的职业生涯中作为怀特输掉的比赛很少。而且,我们的黑开局系统是在尝试下黑的前提下准备的。从来没有准备好处理必须以黑方获胜的情况。我们在第四场比赛中的开局勉强足以保持平局。我们需要一些别的东西,一些风险更大但有获胜机会的东西。穆雷和乔尔决定采用半斯拉夫人防守。他们试图输入尽可能多的可能变化,并在短时间内进行尽可能多的测试。

That night and the next morning, Murray and Joel worked hard. To draw the match, we would have to win the last game with Black. It was a tall order. Garry lost very few games as White in his career. Moreover, our opening system for Black had been prepared under the premise of trying to draw as Black. No opening was ever prepared to handle the case of having to win as Black. The opening we played in game four was barely adequate to try to hold the draw. We needed something else, something that was riskier but held some chance to win. Murray and Joel settled on the idea of playing the Semi-Slav defense. They tried to enter as many possible variations as possible, and did whatever tests they could in the short time available.

第六场比赛遵循半斯拉夫防守的正常路线。然后加里故意改变了他的动作,将深蓝从开篇书中去掉了。我们从来没有让深蓝小号打过这个开局,这一事实立即变得显而易见。深蓝本应将其光方主教发展到b7方格,但在被Garry的换位从开局书中取出后,它改为将主教移至d7方格。深蓝喜欢将象放在开放的对角线上,对角线不再受到棋子的阻碍,并且它不明白有时将象放在封闭的对角线上是可以接受的,只要存在开放对角线的选项。可以通过软件在一定程度上解决这个问题,但完整的解决方案需要修改国际象棋芯片。这场比赛对国际象棋芯片进行了修改,以提高象的处理能力,同样,车也被证明在 1997 年的复赛中至关重要。

Game six followed a normal line for the Semi-Slav defense. Then Garry deliberately transposed his moves to take Deep Blue out of its opening book. The fact that we had never let Deep Blue Jr play this opening became immediately obvious. Deep Blue was supposed to develop its light-squared bishop to the b7 square, but after being taken out of the opening book by Garry’s transposition, it moved the bishop to the d7 square instead. Deep Blue liked to put its bishops on open diagonals, diagonals no longer impeded by pawns, and it did not understand that sometimes it is acceptable to put the bishop on a closed diagonal, as long as the option to open the diagonal existed. It was possible to fix the problem to some extent in software, but a complete solution would require modifying the chess chip. The modifications, implemented as a result of this game, to the chess chip to improve the handling of the bishops and, similarly, the rooks turned out to be critical in the 1997 rematch.

今天,每当我看到第六场剩下的比赛,尤其是在新版深蓝的帮助下,我都会忍不住想:“老深蓝到底是如何在四局比赛后扳平比分的?” 老深蓝做了很多比赛中的站位失误,加里时而哄骗,时而施压,都是造成这些失误的罪魁祸首。深蓝最后一个重大的走位错误是在第 30 步将象移动到 b8,从而将象困在 b8 上,将车困在 a8 上。国际象棋芯片能够检测到主教被困。不幸的是,或者也许幸运的是,从缩短我们的折磨的意义上来说,评估软件对让主教被困的惩罚很少。加里下43手后,我们受够了苦,就辞职去深蓝了。他以4比2的比分赢得了这场比赛。

Today, whenever I look at the remainder of the game six, especially with the help of the newer version of Deep Blue, I cannot help but wonder, “How the hell did the old Deep Blue manage to tie the match after four games?” The old Deep Blue made many positional mistakes in the game, and Garry, sometimes cajoling, sometimes pressuring, had everything to do with causing the mistakes. The last major positional mistake by Deep Blue was to move its bishop to b8 on move thirty, trapping both the bishop on b8 and the rook on a8. The chess chip was able to detect the bishop being trapped. Unfortunately, or perhaps fortunately in the sense of shortening our torment, there was very little penalty levied by the evaluation software for getting the bishop trapped. After Garry’s move 43, we had suffered enough and resigned for Deep Blue. He had won the match by the score of 4 to 2.

在前四场比赛中,深蓝在大部分时间都在走熟悉的道路,并且能够避免出现明显的站位错误。在过去的两场比赛中,Garry将深蓝逼入了未知的领域,探究了它的弱点,利用了它们,将深蓝打得如同一位指挥家一样。珠穆朗玛峰高了一千英尺。

In the first four games, Deep Blue was treading familiar ground and was able to avoid making overt positional mistakes, most of the time. In the last two games, Garry forced Deep Blue into unknown territory, probed for its weaknesses, exploited them, and played Deep Blue as if he were a virtuoso conductor. Mount Everest grew a thousand feet.

1996 年比赛中使用的深蓝 1 号国际象棋芯片。

Deep Blue 1 chess chip, as used in the 1996 match.

    由于完成芯片的巨大压力,我已经好几年没有真正休假了。IBM 不允许员工年复一年地累积假期。我最终在“假期”日去上班。这是我不会再做出的牺牲。

    With the heavy pressure to complete the chip, I had not taken any real vacation days for years by then. IBM did not allow their employees to accumulate vacation days from one year to another. I ended up going to work on “vacation” days. It was a sacrifice that I will not make again.

    自从我们加入 IBM 以来,Jerry 就一直在这个团队中。他并没有真正参与深蓝的技术工作,更多的是提供一般支持,包括零件采购和订单/合同处理。

    Jerry had been with the team ever since we joined IBM. He did not really get involved with the technical work of Deep Blue, but was more of a provider of general support, including parts acquisition and order/contract processing.

    我们从纽约州霍桑搬到了位于纽约州约克敦高地的 IBM 主要研究实验室。

    We had moved from Hawthorne, New York to the main IBM Research lab in Yorktown Heights, New York.

    这是实数。比赛时,CJ Tan 给出的主机 CPU 数量为 32。我们被指示坚持使用数字 32 以避免混淆。该 36 路机器由两个 16-CPU 框架和四个工作站组成。

    This is the real number. At the time of the match, C J Tan gave out the number of host CPUs as 32. We were instructed to stick to the number 32 to avoid confusion. The 36-way machine consisted of two 16-CPU frames plus four workstations.

    软件评估函数为评估项分配权重,而不是像典型的国际象棋程序那样计算评估函数本身。

    The software evaluation function assigned weights for the evaluation terms, instead of computing the evaluation function itself as in typical chess programs.

    换一种说法。每个国际象棋芯片大约相当于一台价值数百万美元的通用超级计算机。特别是,单个国际象棋芯片的国际象棋处理能力实际上比我们用于深蓝的主机 RS/6000 SP 超级计算机更强大。

    Put it in another way. Each chess chip was roughly equivalent to a multi-million-dollar general purpose supercomputer. In particular, a single chess chip was actually more powerful for chess processing than the host RS/6000 SP supercomputer we used for Deep Blue.

    c3 西西里棋实际上在电脑国际象棋比赛中相当常见。大多数程序员没有太多时间来准备他们的入门书籍。

    The c3 Sicilian was actually quite common in computer chess play. Most programmers did not have a lot of time to prepare their opening books.

    与网站管理员最初的预期相反,IBM 比赛网站成为当时最受欢迎的网站,甚至超过了 Superbowl 网站和《体育画报》游泳套装版网站

    Contrary to the web master’s initial expectation, the IBM match web site turned into the most popular web site up to that point in time, exceeding even the Superbowl web site and the swim suite edition web site of Sports Illustrated.

    比赛结束后,Garry 展示了一条可能的平局,需要深蓝立即牺牲一个棋子。深蓝根本看不懂他的画线。

    After the game, Garry showed a possible drawing line that entailed Deep Blue sacrificing a pawn immediately. Deep Blue did not understand his drawing line at all.

   加里对深蓝在第五场比赛中对他的开局选择的立即反应有一个非常奇怪的评论。“你怎么知道我会玩这个?” 他惊呼道。加里很少演奏这句开场白,但他在大师比赛中演奏的却是完全正常的。他似乎认为我们只会准备他通常演奏的开场白。

   Garry had a very odd comment about Deep Blue’s immediate book response to his opening choice in game five. “How could you know that I would play this?” he exclaimed. Garry rarely played this opening line, but what he played was perfectly normal in Grandmaster play. He seemed to believe that we would only prepare for the opening lines that he usually played.

   比赛规则并没有明确允许这种安排。鉴于操作员的棋力相对较弱,包括卡斯帕罗夫团队和比赛仲裁员在内的所有相关各方都事先认为这是唯一可行的解​​决方案。

   The match rules did not explicitly allow this arrangement. All parties involved, including the Kasparov team and the match arbiter, recognized it beforehand as the only practical solution, given the operators’ relatively weak chess playing strength.

 

 

第11章

CHAPTER 11

重组

Retooling


拾起碎片

Picking up the Pieces

从费城回来的路上异常安静。每个人都疲惫不堪,无论是身体上还是精神上。我的心情不是很好,但比开车去看比赛时感觉好多了。失败从来都不是令人愉快的事情,但我心里知道我们会再次向加里发起进攻。

The drive back from Philadelphia was unusually quiet. Everybody was exhausted, both physically and mentally. I was not in the best of moods, but I was feeling much better than during the drive to the match. Losing is never pleasant, but I knew in my heart that we would get another shot at Garry.

我们在费城所取得的成就超出了几乎所有人的期望,包括我们自己。前四局比赛结束后,我们与可能是有史以来最好的国际象棋棋手打成了平手。如果我们有更多时间彻底测试程序,我们可以轻松地赢得第二场比赛。第四局比赛中,加里的结局岌岌可危,他在深蓝的帮助下扳平了比分1。深蓝在前四场比赛后本来可以领先。如果当时我们领先,谁知道会发生什么2?这场比赛引起了巨大的关注。比赛新闻登上了《纽约时报》、《今日美国》甚至可能是美国所有报纸的头版。《今夜秀》、《大卫·莱特曼深夜秀》和《周六夜现场》都在深夜电视节目中恶搞了这场比赛。我当时并不知道,比赛期间全世界记录了超过十亿次“印象” 3 。仅美国市场的同等广告价值就超过 1 亿美元。我们距离击败加里并创造历史已经很近了。活动规模只会越来越大。IBM 绝对不想再打一场。

What we had achieved in Philadelphia exceeded the expectations of almost everybody, including ourselves. After the first four games, we had tied the match with possibly the best chess player ever. We could easily have drawn game two if we had more time to test the program thoroughly. In game four, Garry had a precarious ending that he managed to draw with Deep Blue’s help1. Deep Blue could have been ahead after the first four games. If we had been leading then, who knows what would have happened2? The publicity generated from the match was tremendous; the match news was on the front pages of the New York Times, USA Today, and possibly every newspaper in the United States. The ‘Tonight’ show, the ‘Late Night with David Letterman’ show, and the ‘Saturday Night Live’ show all carried spoofs of the match in their late night TV segments. I did not know it at the time, but over one billion “impressions”3 were recorded worldwide during the match. The equivalent advertising value for the US market alone was worth over $100,000,000. We were within striking distance of beating Garry and making history. The event could only get bigger. There was no way in hell that IBM would not want a rematch.

我休息了一天,然后出于惰性又回到了办公室。穆雷和乔休息了一周。每个人家里都有一个刚出生的婴儿,还有很多事情要做。Joel 与 IBM 的合同到期了,他又回到了在深蓝团队短暂任职之前所熟悉的生活。我确实没有条件工作,只是做一些不需要太多思考的事情。在费城比赛之前,我编写了一个软件评估函数,该函数使用了国际象棋芯片上可用的所有评估功能。编写此代码是为了对硬件评估功能的评估权重自动调整(也称为评估调整)进行实验。该程序测量了从非常浅的搜索中获得的程序的动作与大师在比赛中选择的位置的匹配程度。基本假设是匹配率越大,评价函数越好,国际象棋程序越强。运行实验不需要太多思考,而且由于费城比赛后有充足的机器时间,我花了一整天的时间咀嚼机器周期来运行实验。

I took a day off, and then went back to the office out of inertia. Murray and Joe took a week off. Each had a newborn baby at home, and much catching up to do. Joel’s contract with IBM had expired, and he returned to the life he knew before his short stint on the Deep Blue team. I was really in no condition to work, and was merely doing things that did not require much thinking. Before the Philadelphia match, I wrote a software evaluation function that used all the evaluation features available on the chess chip. This was written to do experiments on automatic adjustment of the evaluation weights, also known as evaluation tuning, for the hardware evaluation features. The program measured how well the program’s moves, obtained from very shallow searches, matched Grandmasters’ choices in positions from their games. The basic assumption was that the greater the matching rate, the better the evaluation function, and the stronger the chess program. Running the experiments did not require a lot of thinking, and since there was ample machine time available after the Philadelphia match, I spent all my days chewing up machine cycles to run the experiments.

这些评估调整实验揭示了一些令人惊讶的事实。特别是,人类大师分配给“开放文件中的车”的权重似乎比国际象棋程序中传统使用的权重低得多。我在费城比赛之前就知道了这一点,但我认为这个结果是我在某处引入的一些软件错误的产物。赛后仔细检查表明,结果是真实的。这并不是说人类大师对“开档车”的权重较低,而是他们在“开档车”何时重要时更加具有辨别力。换句话说,人类大师有意识或无意识地使用的上下文比典型的国际象棋程序要多得多。在这个发现之后,我浏览了整套硬件评估特征,并根据调优程序定位出所有评估权重较低的特征。所有这些功能可能都需要改进。

These experiments on evaluation tuning uncovered some surprising facts. In particular, the weights assigned to “rooks on open files” by human Grandmasters, appeared to be much lower than the ones traditionally used in chess programs. I knew about this before the Philadelphia match, but I thought the result an artifact of some software bugs that I had introduced somewhere. Closer examination after the match indicated that the result was real. It was not that human Grandmasters attached lower weights to “rooks on open files”, but that they were far more discriminating about when “rooks on open files” were important. In other words, there were far more contexts used by the human Grandmasters, either consciously or unconsciously, than by typical chess programs. After this discovery, I went through the entire set of hardware evaluation features, and located all the features that had low evaluation weights according to the tuning program. All these features might be in need of improvement.

默里、乔和我度假回来后开了一次会。我们检查了比赛中出现的问题,并得出了三个结论。首先是深蓝的搜索速度似乎足够。加里在位置上胜过深蓝,但在战术上胜过深蓝。这给我们带来了第二个结论,即与世界冠军相比,深蓝所拥有的国际象棋知识量严重不足。国际象棋芯片对于我们来说太新了,无法在短时间内学习如何有效地使用硬件评估功能。在费城比赛期间,只有大约三分之一的硬件评估功能真正被使用。通过创造性地使用现有的国际象棋硬件,还有很大的改进空间。第三个结论是团队,人类,必须做好更好的准备。当加里在第五场比赛中提出平局提议时,我们还没有做好准备。对于第六局比赛,我们没有准备好以黑方获胜。从心理上和程序上来说,我们还没有为这场重要比赛做好准备。球队一起经历了很多,但我们都没有真正的大型比赛经验。

Murray, Joe, and I had a meeting after they got back from their vacations. We examined what went wrong in the match, and reached three conclusions. The first was that Deep Blue’s search speed appeared to be adequate. Garry was outplaying Deep Blue positionally, but not tactically. This brought us to our second conclusion, which was that Deep Blue was grossly inadequate in the amount of chess knowledge that it had, compared to the World Champion. The chess chip was too new for us to learn how to use the hardware evaluation features effectively in the short time available. During the Philadelphia match, only about one-third of the hardware evaluation features were really in use. There was a lot of room for improvement by creative use of the existing chess hardware. The third conclusion was that the team, the humans, had to be better prepared. We were not prepared when Garry made his draw offer in game five. For game six, we did not have an opening ready to try to win as Black. Psychologically and procedurally, we were not ready for the big match, period. The team had gone through a lot together, but none of us really had any big match experience.

我们不知道什么时候才能再次向加里开枪。乔有些怀疑我们是否还会再有一次机会,但在与默里和我交谈后,他确信我们可能会再试一次,并松了一口气。问题是我们下次可以做些什么来增加胜算。我们可以提高程序的搜索速度,但这可能没有多大帮助。软件评测功能当然还有待改进。我们还可以改进国际象棋硬件,但这只有在复赛前有足够的时间,并且有足够的资金来发布新芯片的情况下才有可能。我们需要尽一切努力说服管理层尽可能晚地安排复赛,并分配足够的预算来发布第二代国际象棋芯片。默里和乔对于经历创造新的国际象棋芯片以及与之配套的所有新软件的前景并不太兴奋。但我说服他们至少要确保我们有足够的时间和资金,以防万一。我们都同意最后一组任务应该是什么。我们需要构建比赛准备工具。加里赢得了费城比赛的部分原因是他比深蓝适应得更快。让深蓝像人类一样具有适应性是很困难的,但我们可以通过作为一个运转良好的团队来提供帮助。在印地 500 赛车比赛中,赛车在绕完几圈后需要在维修站更换轮胎。进站机械师通过团队合作并使用适当的工具,可以在不到十一秒的时间内更换轮胎。我们不需要像进站机制那样快,但我们可以采用相同的方法。我们需要为每个参与者创建一套工具:一套为程序员(我们)提供的工具,一套不同的为大师提供的工具。我们还需要确保团队的所有成员都知道他们的职责是什么。

We had no idea when we would get another shot at Garry. Joe had some doubt whether we would get another chance at all, but after talking with Murray and me, he was convinced and relieved that we would probably have another go. The question was what we could do to increase our odds next time. We could improve the search speed of the program, but that probably would not help much. The software evaluation function could certainly be improved. We could also improve the chess hardware, but that would be possible only if there was enough time before the rematch, and there was enough money allocated to make a new chip release. We would need to do whatever we could to convince the management to schedule the rematch as late as possible, and to allocate an adequate budget to allow for the release of a second-generation chess chip. Murray and Joe were not too thrilled about the prospect of going through the gyration of creating a new chess chip, and all the new software to go with it. But I persuaded them to at least make sure that we had adequate time and funding just in case. We all agreed on what the last set of tasks should be. We needed to build tools for match preparation. Garry won the Philadelphia match in part because he could adapt faster than Deep Blue. It was difficult to make Deep Blue as adaptive as a human being, but we could help by working as a well-oiled team. In Indy 500 car races, after going around the laps several times the cars need to get their tires replaced at the pit stop. The pit-stop mechanics, by working as a team, and with the proper tools, can replace the tires in less than eleven seconds. We did not need to be as fast as pit-stop mechanics, but we could adopt the same approach. We needed to create sets of tools for everyone involved: a set of tools for the programmers (us), and a different set for the Grandmasters. We also needed to make sure that all members of the team knew what their responsibilities were.

然而,在创建新工具之前,我与默里和乔就是否应该有新芯片以及新芯片中应包含哪些内容进行了辩论。这场辩论持续了几个月。

Before the new tools were created, however, I got into a debate with Murray and Joe on whether there should be a new chip, and then on what should go into the new chip. This debate lasted several months.

再过六个月

Another Six Months

不知道为什么,经过长途海外旅行后,我似乎能够集中注意力。1985年,在我去台湾旅行之后,我能够在六个月内投入完成移动发电机芯片所需的能量。1996 年,费城比赛结束后大约一个月,我被邀请在 IBM 东京研究实验室发表有关深蓝的演讲。IBM 研究中心正在举办自己的 50 周年庆典,作为庆祝活动的一部分,东京实验室正在举办为期一天的并行处理会议。我的演讲是会议的其中一场会议。在这次东京之旅之后的六个月里,我比我一生中任何时候都更加努力地工作,甚至比我设计第一个国际象棋芯片时还要努力。

For some unknown reason, I seem to be able to concentrate well after a long overseas trip. In 1985, after my trip to Taiwan, I was able to pour in the energy needed to finish the move generator chip, in six months. In 1996, about a month after the Philadelphia match, I was asked to give a talk about Deep Blue at the IBM Tokyo Research Lab. IBM Research was having its own fiftieth anniversary celebration, and the Tokyo Lab was hosting a one-day conference on parallel processing as part of the celebration. My talk was one of the sessions in the conference. For the next six months after this Tokyo trip, I worked harder than I ever had in my entire life, harder even than when I designed my first chess chip.

1985 年,国际象棋走棋生成器是否会被制造出来是毫无疑问的,因为 MOSIS 服务实际上是免费的。1996年,我开始研究新版本的深蓝国际象棋芯片,但不知道是否有资金来建造它。除了开发新芯片之外,我还必须扮演倡导者的角色,首先说服团队的其他成员,然后说服我们的经理,新芯片是可取的和必要的。

In 1985, there was no question over whether the chess move generator would be fabricated, as the MOSIS service was effectively free. In 1996, I began work on the new version of the Deep Blue chess chip without knowing whether there would be money to build it. Besides working on the new chip, I also had to wear the hat of an advocate, first persuading the rest of the team, and then our manager, that the new chip was desirable and necessary.

考虑到所有的不确定性,我在芯片设计和试图说服所有人方面都采取了分阶段的方法。新功能分阶段添加到国际象棋芯片中,在此过程中始终保持新的“工作”版本。这意味着我可以随时停止新设计,并拥有一些原来的东西准备在很短的时间内进行制造。芯片设计过程比较简单,我只要投入工作就可以了;成为倡导者的任务要困难得多。

Given all the uncertainty, I took a staged approach, both in the chip design and in my attempt to persuade everyone. New functions were added to the chess chip in stages, always maintaining a new “working” version in the process. This meant I could stop the new design at any point, and have something that was ready for fabrication on very short notice. The chip design process was relatively straightforward, I just had to put in the work; the task of being an advocate was much harder.

最初,拨出资金只是为了制造更多 1996 年比赛中使用的棋子。这需要一些努力,但我设法说服团队的其他成员,在进行另一次芯片运行之前,我们至少应该在国际象棋芯片上添加一个重复检测器。默里、乔和我观察到小深蓝下的奇怪棋步似乎是由于国际象棋硬件中缺乏重复检测而产生的“地平线效应”造成的。Deep Blue Jr 有时会尝试重复它喜欢的位置,方法是将位置推入搜索树超出其“重复检测范围”的部分,这是具有重复检测功能的软件搜索树的“范围”。虽然这种行为肯定会产生有害影响,但在没有明确证据表明有必要的情况下,默里和乔都不愿意支持这一添加。事实证明,香港发生的灾难正是我所需要的说服力。世界国际象棋比赛的平局原本是可以避免的。在我们输给弗里茨的过程中,许多其他事情同时出现了问题,但是有了硬件重复检测器,机器就可以首先避免出现问题。可以想象,《Deep Thought II》可能会以满分的成绩获得香港赛的冠军;换句话说,其性能将高出 200 多个评级点。最后的论点引起了默里和乔的注意,他们同意支持这个想法。这是在我访问东京之前发生的。

Originally, money had been set aside only for fabricating more of the same chess chips that were used in the 1996 match. It took some doing, but I managed to convince the rest of the team that we should at least add a repetition detector to the chess chip before going through another chip run. Murray, Joe and I, had observed Deep Blue Jr playing strange moves that appeared to result from the “horizon effect” related to the lack of repetition detection in the chess hardware. Deep Blue Jr would sometimes try to repeat a position that it liked by pushing the position into the part of the search tree beyond its “repetition detection horizon”, which was the “horizon” for the software search tree that did have repetition detection. While this behavior had a definitely detrimental effect, neither Murray nor Joe was willing to endorse the addition without clear evidence that it was necessary. The disasters that we had in Hong Kong turned out to be the persuasion that I needed. The draw given up to WChess could have been avoided. Many other things went wrong simultaneously in our loss to Fritz, but with a hardware repetition detector, the machine would have avoided the problem line in the first place. It was conceivable that Deep Thought II might have won the Hong Kong event with a perfect score; in other words, its performance would have been more than 200 rating points higher. This final argument got both Murray’s and Joe’s attention, and they agreed to endorse the idea. This took place before my visit to Tokyo.

在费城比赛之前的一段时间,我意识到有一种更好的方法来实现评估功能,可以显着减少芯片面积。在往返东京的航班中,我意识到这一实施也使得评估功能变得更加全面。问题是,这将需要对大约三分之二的国际象棋筹码进行大规模更改。这种改变是非常可取的,但风险也很大。我相信如果我们想击败加里,就必须进行这种改变,而且我真的很想赢。

Some time before the Philadelphia match, I realized that there was a better way to implement the evaluation function that would reduce the chip area significantly. During the flights to and from Tokyo, it dawned on me that this implementation also made it possible for the evaluation function to be far more comprehensive. The problem was that this would entail a wholesale change to roughly two-thirds of the chess chip. The change was highly desirable, but it was also very risky. I believed that this change had to be done if we wanted to beat Garry, and I really wanted to win.

我的东京之行结束后,我们与 CJ 举行了一次团队会议;默里和乔帮助我说服他探测器必须包含在新芯片中。但三人都对创建全新评估函数的想法犹豫不决。最后,大家同意我设计两个版本的国际象棋芯片,都带有硬件重复检测器,但一个具有旧的评估函数,一个具有新的评估函数。我的首要任务是使用旧的评估函数创建版本。我可以花时间研究具有新评估功能的版本,但我不会从穆雷或乔那里得到任何帮助——他们有两套比赛准备工具需要创建。在这两种新芯片中,只会制造一种。为了获得新的评估函数,我必须自己完成重复检测器的设计新的评估函数,以及对整个芯片进行彻底的验证。我有四个月的时间来决定使用哪个版本。我同意额外的工作量,因为它提供了另一个机会来创造一个足够强大的国际象棋机器来击败世界国际象棋冠军,并且我将尽我所能来实现这一目标。

We had a team meeting with C J after my trip to Tokyo; Murray and Joe helped me persuade him that the detector had to be included in the new chip. But all three balked at the idea of creating a brand-new evaluation function. In the end, it was agreed that I would design two versions of the chess chip, both with the hardware repetition detector, but one with the old evaluation function, and one with the new. My first priority was to create the version with the old evaluation function. I could spend time on the version with the new evaluation function, but I would not be getting any help from Murray or Joe—they had two sets of match preparation tools to create. Of the two new chips, only one would be fabricated. To get the new evaluation function, I would have to complete the design of the repetition detector, and the new evaluation function, as well as having the entire chip thoroughly verified, all by myself. I had four months until we would decide-which version to use. I agreed to the extra workload as it provided another chance to create a chess machine that was strong enough to beat the World Chess Champion, and I was going to do everything within my power to make that happen.

第一个月相对容易,主要花在设计和验证重复检测器上。到第一个月月底,我就准备好了第一个新版本。距离我真正想要的更新版本还有三个月的时间。根据第一款深蓝国际象棋芯片的经验,穆雷和乔推测,即使有他们的帮助,至少也需要一年的时间。有几个原因让我相信我能成功。首先,第一个深蓝国际象棋芯片的设计留下了相当多的软件工具,我可以利用旧的工作来节省一些时间。其次,我得到了休息,如果有必要,我可以投入更长的时间。第三,作为唯一的设计师,我对设计内容有全面的了解,并且我发现问题和解决问题的周期会短得多。最后,最重要的是,我想赢。

The first month was relatively easy, and mostly spent designing and verifying the repetition detector. By the end of the first month, I had the first new version ready. There were three months left for the even newer version that I really wanted. Based on the experience from the first Deep Blue chess chip, Murray and Joe surmised that it would have taken at least a solid year even with their help. There were several reasons why I believed that I could pull it off. First, there were quite a number of software tools left over from the design of the first Deep Blue chess chip, and I could leverage the old work to save some time. Second, I was rested and I could pour in longer hours if necessary. Third, being the sole designer, I had the full picture of what was in the design, and I would have much shorter cycles of finding problems and fixing them. Finally, above all, I wanted to win.

第二个月用于重新设计新的评估功能。由于这是一次大规模的改变,我无法判断是否有任何东西起作用,直到新设计部分完成,大约进入第二个月的两周。我很高兴新电路通过了我的初步测试——新设计原则上有效。现在最困难的部分来了。必须创建软件模型,并且设计和软件模型必须相互匹配。软件模型必须根据设计意图创建,而不是根据设计内容创建,以避免任何设计错误永久存在。这个软件模型几乎是全新的,因为新的评估功能与旧的有很大不同。使用了旧芯片软件模型的各个部分,但主要作为构建块。接下来的一个半月,直到第三个月末,我每周工作时间接近 100 小时。没有周末。每天晚上,当我深夜从实验室开车回家时,我都会反复告诉自己,“我想赢。” 回家后,我会检查电路仿真以确保它仍在运行。如果我在夜间醒来,我会再次检查模拟,如有必要,在重新启动模拟之前修复我的软件模型或其他内容。

The second month was spent redesigning the new evaluation function. Since it was a wholesale change, I could not tell whether anything was working until the new design was partially completed, about two weeks into the second month. I was thrilled that the new circuit passed my initial test—the new design worked in principle. Now came the hard part. A software model had to be created, and the design and the software model had to match each other. The software model had to be created based on the design intent, not based on what was in the design, to avoid perpetuating any design error. This software model was almost completely new, as the new evaluation function was quite different from the old one. Pieces from the software model for the old chip were used, but mainly as building blocks. For the next month and half, until the end of the third month, I was working close to 100 hours per week. There were no weekends. Every night when I drove home from the lab late at night, I would tell myself repeatedly, “I want to win.” After I got home, I would check my circuit simulation to make sure it was still running. If I woke up during the night, I would check the simulation again, and if necessary, fix my software model or whatever before restarting the simulation.

到了第三个月末,我对新的评价函数进行了合理的验证。当然,如果没有旧芯片留下的软件工具,这是不可能实现的。我开始变得雄心勃勃。软件模型和硬件匹配得很好,我不再需要半夜醒来检查东西。由于我比计划提前了,我想我不妨寻找其他可以改进的地方。在第一个深蓝国际象棋芯片制造出来之前,我最初的设计有一个复杂的国际象棋走法生成器,可以“修剪”掉那些已证明无用的国际象棋走法,理论上可以有效地将国际象棋芯片的速度提高三到五倍。复杂的着法生成器的计划在第一款深蓝国际象棋芯片中被放弃,因为所涉及的计算的正确性很难验证。没有足够的时间来完成我之前计划的原始移动生成器。事情太复杂了。但如果我愿意放弃对如何修剪国际象棋动作的“正确性”要求,我可以做一些中等复杂度的事情。我找到了一个令我满意的修剪方案,继续实施新的移动生成器,而没有告诉团队的其他成员或我们的经理。为了安全起见,我向新的移动生成器添加了一种模式,使其能够以与旧的移动生成器相同的方式运行。最终,在完成并测试后,我确实通知了默里和乔。他们并不兴奋。

By the end of the third month, I had the new evaluation function reasonably verified. Of course, this would have not been possible without the software tools left over from the old chip. I was starting to get ambitious. The software model, and the hardware were matching quite well, and I no longer had to wake up in the middle of the night to check things. Since I was ahead of schedule, I figured that I might as well look for whatever other improvements I could make. Before the first Deep Blue chess chip was fabricated, my original design had a complicated chess move generator that could “prune” away provably useless chess moves, which theoretically could effectively speed up the chess chip by three to five times. The plan for the complicated move generator was dropped in the first Deep Blue chess chip, as the correctness of the computation involved was quite difficult to verify. There was not enough time to do the original move generator that I planned earlier. It was far too complicated. But I could do something intermediate in complexity if I was willing to drop the “correctness” requirements on how I pruned away the chess moves. I found a pruning scheme that I was happy with, went ahead and implemented the new move generator, without telling either the rest of the team or our manager. To be on the safe side, I added to the new move generator a mode that allowed it to behave the same way as the old move generator. Eventually, after it was completed and tested, I did inform Murray and Joe. They were not thrilled.

新的移动生成器以及其他领域的一些改进是在第四个月设计和验证的。随着第四个月末的临近,默里和乔清楚地意识到我的新设计已经准备好了,我们有意识地决定根据新设计定制所有新软件开发。新的评估函数提供了如此多强大的新功能,默里和乔都认为进行转换是值得的。即使新芯片存在一些小问题,它所提供的优势也是压倒性的。

The new move generator, along with some improvements in other areas, was designed and verified in the fourth month. As the end of the fourth month drew near, it became clear to Murray and Joe that my new design was ready and we made the conscious decision to tailor all the new software development to the new design. The new evaluation function offered so many powerful new features that both Murray and Joe decided that making the switch would be worthwhile. Even if there were some minor problems with the new chip, the advantages provided by it were overwhelming.

事实证明,距离设计必须交付给 VLSI Technology 还剩两个月的时间。我利用这段时间进行了一些小的渐进式改变。

It turned out that there were two more months to go before the design had to be delivered to VLSI Technology. I used the time to make small incremental changes.

同时,在我们决定使用新的评估功能后,Murray采用了它的软件模型,添加了图形界面,并创建了一个强大的分析工具。这个新软件成为程序开发和复赛期间的重要工具。乔尔大约在这个时候重新加入了该小组,并被介绍使用该工具。很可能,新国际象棋芯片上最重要的新评估功能是我们与他互动的结果。

Meanwhile, after we made the decision to use the new evaluation function, Murray took its software model, added a graphical interface, and created a powerful analysis tool. This new piece of software became an important tool during the program development, and during the rematch itself. Joel rejoined the group at about this time, and was introduced to the tool. Quite likely, the most important new evaluation features on the new chess chip were added as a result of our interactions with him.

有一天,当我正在研究刚刚添加的一项新评估功能时,乔·霍恩 (Joe Hoane) 非常兴奋地来到我的办公室。他说,“有一些东西你应该看看”,然后带我去了乔尔的办公室。穆雷已经在那里了。显然,乔尔一直在与穆雷和乔谈论深蓝小号在乔尔刚刚对阵的比赛中遇到的一些问题。他以前也遇到过类似的问题,但直到他得到了评估函数的新分析工具后才查明确切的根源。

One day, while I was going over a new evaluation feature that I just added, a very excited Joe Hoane came into my office. He said, “There is something that you should see”, and took me to Joel’s office. Murray was already there. Apparently, Joel had been talking with Murray and Joe about some problems that Deep Blue Jr had in a game Joel had just played against it. He had seen similar problems before, but could not pinpoint the exact source until he got the new analysis tool for the evaluation function.

Deep Blue Jr 以及大多数其他现代国际象棋程序都知道将车放在开放的文件上(不再受到棋子阻碍的文件上)的重要性,以便车能够发挥最大的影响力。问题是,当时,与我所知道的任何其他国际象棋程序一样,Deep Blue Jr 并不知道将车放置在可以选择“打开”文件的文件上是有利可图。也就是说,通常通过捕获棋子来移除阻碍车的棋子。在乔尔刚刚打的那场比赛中,深蓝Jr可以将车放在这样一个潜在的空档上,间接施加长期压力。相反,深蓝 Jr 过早地打开了该文件,然后将其车放在打开的文件上。但这一行动让乔尔有机会用自己的车挑战空档,然后沿着空档进行交易。随着所有的车都消失了,深蓝的位置优势消失得无影无踪。乔尔建议,打这个位置的正确方法是将车堆放在可能打开的文件上,然后等待最合适的时机再打开它。同样,同样的概念也体现在象与对角线之间的关系上。第一场比赛的第六局是深蓝错误对待对角线上主教的典型例子。经过一番讨论,我们提出了识别相关特征的规则。当天下午,我就安装并测试了新功能,这是一个相对容易的修改。这一新功能出现在 1997 年复赛的关键第二场比赛中,当时深蓝因潜在的开档而将加里的重型棋子压在了后面。如果没有这个新功能,深蓝就不太可能获胜。对于一个下午的工作来说还不错。

Deep Blue Jr, as well as most of the other modern chess programs, understands the importance of putting rooks on open files, files that are no longer impeded by pawns, so that the rooks can exert the greatest influence. The problem was that, at the time, along with any other chess programs that I knew of, Deep Blue Jr did not know that rooks could be profitably placed on files where it had the option of “opening up” the files. That is, the option of removing the pawns impeding the rooks, usually by making a pawn capture. In the game that Joel just played, Deep Blue Jr could have piled up its rooks on such a potentially open file and exerted long term pressure indirectly. Instead, Deep Blue Jr opened up the file prematurely and then placed its rooks on the resultant open file. But this action gave Joel the chance to challenge the open file with his own rooks, which then were traded off along the open file. With all the rooks gone, Deep Blue Jr’s positional advantage dissipated into thin air. Joel suggested that the right way to play the position was to just pile up the rooks on the potentially open file, and wait for the most opportune moment before opening it. Similarly, the same concept also shows up in the relationship between the bishops and the diagonals. Game six in the first match was a prime example of Deep Blue’s incorrect treatment of bishops on diagonals. After some discussion, we came up with rules for recognizing the relevant features. That very afternoon, I put in and tested the new features which was a relatively easy modification. This new feature showed up in the critical game two of the 1997 rematch when Deep Blue pinned down Garry’s heavy pieces to his back rank as a result of a potentially open file. Without the new feature, it was very unlikely that Deep Blue would have won. Not bad for an afternoon’s work.

在芯片设计的最后两个月中还添加了其他主要的新评估功能。第一个深蓝国际象棋芯片对国王的安全性进行了相当全面的评估,但 1996 年的比赛提供了充分的证据表明还需要做更多的工作。在芯片设计的最后两个月,借助评估函数分析工具,对王安全评估逻辑进行了大规模的重新设计。这种重新设计对深蓝在复赛的第五场和第六场比赛中的表现产生了积极的影响。毫无疑问,它也或多或少地改变了其他比赛的进程。

Other major new evaluation features were also added during the last two months of chip design. The first Deep Blue chess chip had a fairly comprehensive evaluation for king safety, but the 1996 match provided ample evidence that still more work was needed. A massive redesign of the king safety evaluation logic took place during the last two months of chip design, with the help of the evaluation function analysis tool. This redesign positively influenced Deep Blue’s play in games five and six of the rematch. Undoubtedly, it also altered the course of the other games at one point or another, for better or for worse.

幻影女王

The Phantom Queen

新国际象棋芯片的网表于 1997 年 9 月发送给 VLSI Technology 签约的设计公司进行物理设计。新芯片的制造于 1996 年 12 月下旬开始。新国际象棋芯片的印刷电路板是并行设计的,并在 1997 年 2 月新芯片运回之前投入使用,距离复赛还不到三个月。经过与加里的长期谈判,复赛定于五月举行。复赛的准备时间只比第一场比赛的时间稍微好一点。然而,这次软件的状况要好得多。我们已经有了基于旧芯片的工作软件,并且在我们关注新芯片之前,我们已经有了状态良好的新软件评估功能。

The netlist for the new chess chip was sent out in September 1997 to a design firm contracted by VLSI Technology to do the physical design. Fabrication for the new chips started in late December 1996. The printed circuit boards for the new chess chips were designed in parallel, and became available before the new chips were shipped back in February 1997, less than three months before the rematch. After long negotiations with Garry, the rematch had been scheduled to take place in May. The amount of time left for preparation for the rematch was only slightly better than the time left for the first match. However, the software was in much better shape this time. We already had working software based on the old chips, and we had the new software evaluation function in pretty good shape before we ever laid eyes on the new chips.

新国际象棋硬件与新软件的最初集成进展顺利,但随后我们开始遇到障碍。Joe 负责软件集成的大部分工作,他有条不紊地完成了集成。他从新旧软件的共同功能开始,然后一次添加一个新功能。他首先将旧软件移植到新的国际象棋硬件上,进展顺利。然后,他简化了国际象棋硬件的软件接口,因为新的国际象棋硬件允许更简单的接口。这也很顺利。乔采取的下一步是激活硬件重复检测器,但他很快就遇到了问题。

The initial integration of the new chess hardware with the new software went smoothly, but then we started to hit snags. Joe, who was doing the bulk of work for the software integration, did the integration methodically. He started with the common features of the old and the new software, then added new features one at a time. He first ported the old software to the new chess hardware, and it went fine. He then simplified the software interface to the chess hardware, as the new chess hardware allowed a simpler interface. This also went well. The next step Joe took was to activate the hardware repetition detector where he soon ran into problems.

我用来模拟新的国际象棋芯片的设计软件有些限制。理想情况下,我们希望能够在模拟国际象棋芯片上运行国际象棋软件,这样在芯片制造之前就可以发现国际象棋软件与芯片集成的任何问题。我们面临两个障碍。首先,没有简单的方法可以做到这一点。设计软件没有一个简单的机制让我们可以直接从国际象棋软件驱动模拟国际象棋芯片。其次,模拟的国际象棋芯片比真实的国际象棋芯片慢一百万倍以上,而真实芯片上一分钟的运行时间将花费我们两年多的模拟时间。1996年和1997年的比赛之间只有一年多一点的时间。通常,在真实芯片上运行至少几个小时才能发现与系统集成相关的所有问题。因此,新的国际象棋芯片在提交制造之前从未真正得到我完全满意的验证。更先进的设计软件已经可用,但没有时间将设计转换为与较新的设计软件一起运行。此外,购买更新的设计软件需要花费很多钱。

The design software that I used to simulate the new chess chip was somewhat limited. Ideally, we would have liked to be able to run the chess software on top of the simulated chess chip, so that any problems with integrating the chess software with the chip could be discovered before the chips were fabricated. There were two obstacles that faced us. First, there was no easy way to do it. The design software did not have a simple mechanism that allowed us to drive the simulated chess chip directly from the chess software. Second, the simulated chess chip was over a million times slower than the real chess chip, and one minute of running time on the real chip would have cost us over two years of simulation time. There was only just slightly over one year of time between the 1996 and the 1997 match. Usually, it would take at least several hours of running time on the real chip to uncover all the problems associated with the system integration. So the new chess chip was never really verified to my complete satisfaction before it was submitted for fabrication. More advanced design software was already available, but there was no time to convert the design to run with the newer design software. Besides, it cost big bucks to get the newer design software.

硬件重复检测器经过了彻底的测试,至少我是这么认为的。

The hardware repetition detector was thoroughly tested, or so I thought.

1992年,当我将国际象棋软件的控制权转移给乔时,我以某种方式编写了国际象棋程序。我对硬件重复检测器的所有测试都是基于 Deep Thought II 软件与国际象棋硬件的接口方式。在接下来的四年里,乔对软件进行了重大更改和改进。到 1996 年初,他的新程序仅用于搜索部分就已增长到超过 100,000 行代码,大约是我 1992 年给他的代码大小的 20 倍。不用说,乔以更多的方式访问国际象棋硬件。复杂的时装,并使用了全套所有可能的操作。我根据我编写的国际象棋程序的方式模拟并测试了硬件重复检测器,但是使用 Joe 的代码,它无法正常工作。一些芯片操作没有在我的模拟中测试,但被 Joe 在他的代码中使用,具有向硬件重复检测器提供错误数据的副作用。

In 1992, when I transferred the control of the chess software to Joe, I wrote the chess program in a certain way. All my testing of the hardware repetition detector was based on how I interfaced the Deep Thought II software with the chess hardware. Joe had changed and improved the software substantially in the intervening four years. By early 1996, his new program for the search part alone had grown to well over 100,000 lines of code, or about 20 times the size of the code that I gave him in 1992. Needless to say, Joe accessed the chess hardware in a more complicated fashion, and used the full set of all possible operations. I simulated and tested the hardware repetition detector based on how I wrote the chess program, but with Joe’s code, it did not work properly. Some of the chip operations that were not tested in my simulation, but were used by Joe in his code, had the side effect of giving false data to the hardware repetition detector.

这个问题并不致命,而且解决方案也比较容易,只要你不用写代码。他修改了国际象棋程序,以使用与硬件重复检测器一起正常工作的较小的芯片操作集。这花了我们大约一周的时间,但乔最终成功了。

The problem was not fatal, and the solution was relatively easy, as long as you didn’t have to write the code. He modified the chess program to use the smaller set of chip operations that worked properly with the hardware repetition detector. This cost us about a week but Joe got it working in the end.

然而,他发现的下一个问题却没有令人愉快的解决方案。关于新芯片的一大问题是硬件搜索修剪的效果如何。Joe启用硬件搜索修剪后,早期结果相当不错。修剪似乎将搜索速度提高了 5 到 10 倍,或多或少符合我的估计。然后乔发现了一些奇怪的事情。在他正在测试的一个位置中,搜索完成后,芯片内部棋盘的方格(左下角)上出现了一个额外的白皇后。无论使用哪种芯片,该问题都会重复出现。几乎可以肯定我在某个地方引入了逻辑设计错误。迫在眉睫的问题是什么以及在哪里?

The next problem that he uncovered, however, had no pleasant solution. One of the big questions about the new chip was how well the hardware search pruning really worked. After Joe enabled the hardware search pruning, the early results were fairly good. The pruning appeared to speed up the search by a factor of 5 to 10, more or less in line with my estimate. Then Joe found something peculiar. In one of the positions that he was testing, an extra White queen showed up on the al square (lower left corner) of the chip’s internal chessboard after the search was completed. The problem was repeatable no matter which chip was used. It was almost certain that I had introduced a logic design error somewhere. The immediate questions were what and where?

乔在三层搜索中找到了幻影女王出现的位置。模拟大约需要一天时间才能重现该错误。我最不喜欢的国际象棋规则又回来困扰着我。幻影女王只有在可以旁观的情况下才会出现。所以这又是顺便了。在某些罕见的情况下,al 上的白皇后或 a8 上的黑皇后可能会凭空出现。然而,如果我们禁用硬件搜索修剪,幽灵女王就不会出现。

Joe located a position where the phantom queen showed up in a three-ply search. It took about one day of simulation to replicate the error. My least favorite chess rule came back to haunt me. The phantom queen only showed up when en passant was possible in the position. So it was en passant yet again. Under certain rare conditions, either a White queen on al, or a Black queen on a8 could materialize out of thin air. The phantom queens, however, would not show up if we disabled the hardware search pruning.

我们有三种方法来解决幻影女王问题。

We had three ways to solve the phantom queen problem.

最简单的方法是禁用硬件搜索修剪,放弃修剪带来的五到十倍的加速。作为让某些东西立即运行的短期解决方案,这已经足够了。

The simplest was to just disable the hardware search pruning, giving up the five- to ten-fold speedup from the pruning. This was adequate as a short-term solution to get something running immediately.

我更喜欢第二种方法,但在时间上不可行。这是一个硬件解决方案。新的国际象棋芯片设计为可通过 FPGA(现场可编程门阵列)进行外部控制,成本约为 20 美元。外部 FPGA 可以监控硬件搜索并在幽灵皇后出现时将其移除。这种方法还使我们可以选择在 FPGA 内部添加额外的评估功能,而无需对国际象棋芯片进行昂贵的重新设计。硬件方法的另一个优点是软件基本保持不变。距离复赛还有两个多月的时间。这个硬件解决方案需要一批新的印刷电路板,而我们的常规印刷电路板供应商表示,他们最多可以在复赛前两周左右供应这些电路板。即使对我来说,两周也太紧了。我们也许能够找到一家可以更快供应电路板的新供应商,但这个提议风险太大。

The second approach, which I preferred, was not feasible time-wise. It was a hardware solution. The new chess chip was designed to be controllable externally with an FPGA (Field Programmable Gate Array), costing around $20. The external FPGA could monitor the hardware search and remove the phantom queens as they showed up. This approach would also give us the option of adding additional evaluation features inside the FPGA, without going through the costly redesign of the chess chip. Another advantage of the hardware approach was that the software remained essentially the same. There were slightly over two months left before the rematch. This hardware solution required a new batch of printed circuit boards, and our regular printed circuit board supplier said that at best they could supply the boards about two weeks before the rematch. Two weeks was too tight even for my taste. We might have been able to find a new supplier who could supply the boards faster, but it was too risky a proposition.

第三种方法需要大量的软件更改,我认为我们最好坚持使用第一种方法中最简单的解决方案。我对在比赛前做出这样的改变感到不​​舒服。这种方法是基于这样的观察:每当硬件搜索后幽灵女王出现时,位置评估几乎总是略有变化。该提案是在搜索之前启用硬件搜索剪枝作为默认值,并且当在默认搜索之后观察到评估发生变化时,删除幻影皇后,禁用硬件剪枝,并重新搜索位置。通常情况下,需要重新搜索的棋子位置不到十分之一。因此,有效速度比设计速度慢大约百分之三十到一百,但比最简单的软件修复提供的速度快得多。这种方法存在两个正确性问题。首先,当硬件搜索后幽灵女王确实出现时,不能保证会观察到位置评估的变化。其次,在适当的条件下,幻影女王可能会在搜索过程中出现,但在完成搜索之前就被覆盖,从而导致错误的搜索结果,但没有任何故障迹象。换句话说,如果采用第三种方法,由于幽灵女王问题,搜索不能完全可信。这是一个智力上有趣的解决方案,这可能就是乔决定花大约一个月的时间让它发挥作用的原因。

The third approach required massive software changes, and I was of the opinion that we were better off sticking with the simplest solution of the first approach. I was not comfortable with making such changes right before the match. This approach was based on the observation that whenever the phantom queen appeared after the hardware search, the positional evaluation almost always changed slightly. The proposal was to enable the hardware search pruning before the search as the default, and when a change in evaluation was observed after the default search, remove the phantom queen(s), disable the hardware pruning, and re-search the position. Normally, less than one in ten positions given to the chess chips needed to be re-searched. The effective speed was thus about thirty to one-hundred percent slower than the design speed, but much faster than the speed provided by the simplest software fix. There were two correctness problems with this approach. First, there was no guarantee that a positional evaluation change would be observed when a phantom queen did show up after hardware search. Second, the phantom queen, given the right conditions, could show up during the search, but then get overwritten before completing the search, and thus cause a false search result yet without any indication of the fault. In other words, the search could not be completely trusted because of the phantom queen problem, if the third approach was adopted. It was an intellectually interesting solution, which was probably why Joe decided to spend about a month getting it to work.

幽灵女王问题之后没有发现其他硬件错误。这个问题最终花费了我们大约一个月的时间,并且可能导致软件变得不太稳定。最后一个已知的由幽灵女王问题引起的软件错误直到 1997 年复赛前两周才得到修复。最后,一切都很顺利,但有一段时间事情看起来很糟糕。

There were no other hardware errors uncovered after the phantom queen problem. The problem ended up costing us about one month of Joe’s time, and probably made the software less stable as a result. The last known software bug caused by the phantom queen problem was not fixed until two weeks before the 1997 rematch. In the end, everything worked out all right, but things did look bad for a while.

质量先于数量

Quality before Quantity

1996年的比赛中,加里祝贺球队成功地“把数量转化为质量”。他当时没有意识到,深蓝从来不只注重数量。特别是在第三局抽签中,国际象棋芯片中的一些新的评估功能显着地发挥了作用。有些“不人道”的举动根本不是深度搜索的结果,而是深蓝评估加里的车已接近被困的结果。

During the 1996 match, Garry congratulated the team for succeeding in “converting quantity into quality”. What he did not realize at the time was that Deep Blue never had just quantity. In particular, in the drawn game three, some of the new evaluation features in the chess chip came prominently into play. Some of the “inhuman” moves were not the result of deep searches at all, but of Deep Blue evaluating Garry’s rook getting close to being trapped.

1997年的复赛,我们备战的重点是提高深蓝棋艺的基础素质。这一次,我们希望质量先于数量。

For the 1997 rematch, the main emphasis of our match preparation was to improve the underlying quality of Deep Blue’s chess knowledge. This time, we wanted the quality to go in before the quantity.

创造新的国际象棋筹码只是我们追求质量的第一步。

Creating the new chess chip was just the first step in our quest for quality.

进入1996年的比赛时,深蓝只存在了两周。那是一个婴儿。这是一个非常强大、非常有才华的婴儿。否则它不可能在第一场比赛中就击败加里。但 1996 年比赛的结果也痛苦地表明,婴儿需要接受交易技巧。到了 1996 年 9 月,复赛已经确定,乔尔·本杰明被邀请回到 IBM 研究中心,担任一所国际象棋学校的校长,该学校只有一个学生,即“深蓝”。(实际上,大多数时候,学生实际上是 Deep Blue Jr,因为我们无法完全访问 Deep Blue 运行的 IBM RS6000 SP 超级计算机。)

Going into the 1996 match, Deep Blue had only existed for two weeks. It was a baby. It was a very powerful and very talented baby; otherwise it could not have possibly beaten Garry in its very first game. But the result of the 1996 match also made it painfully clear that the baby needed to be taught the tricks of the trade. By September 1996, it was certain that the rematch would take place, and Joel Benjamin was invited back to IBM Research to be the headmaster of a chess school with only a single pupil, namely, Deep Blue. (In reality, most of the time, the pupil was actually Deep Blue Jr, since we did not have full access to the IBM RS6000 SP supercomputer that Deep Blue ran on.)

典型的一天的学校生活是从乔尔和小深蓝下“收回”国际象棋开始的;有时这种情况会持续到下午早些时候。即使在早期,深蓝Jr可能已经达到了超级宗师(2600+国际评级)的实力,所以当乔尔的位置开始对他不利时,他允许自己收回自己的动作。通过反复回到自己还有机会的位置,他发现了很多给深蓝带来严重问题的位置。在“收回”国际象棋课程之后,他会告诉我们小深蓝所犯的错误。然后我们要么对小深蓝进行更改,要么在某些情况下确保问题在新的中得到解决。国际象棋筹码。第二天,整个过程又开始了,通常会有一些惊喜等待着乔尔。改良后的深蓝Jr经常会出昨天不会出的招式。

A typical school day started with Joel playing “take-back” chess with Deep Blue Jr; sometimes this continued in the early afternoon. Deep Blue Jr was probably already of Super Grandmaster (2600+ international rating) strength even in the early days, so Joel allowed himself to take back his own moves when the position started to look bad for him. By going back repeatedly to positions where he still had chances, he found lots of positions that caused Deep Blue Jr serious problems. After the “take-back” chess sessions, he would tell us about the mistakes made by Deep Blue Jr. We would then either make changes to Deep Blue Jr or, in some circumstances, make sure that the problems were dealt with in the new chess chip. The next day, the whole process started again, usually with some surprises waiting for Joel. The newly improved Deep Blue Jr would frequently play moves that it would not have played the previous day.

每天的变化大部分是在它的评价函数中。

Most of the daily changes were in its evaluation function.

到目前为止,深蓝团队就像一个三驾马车,每个团队成员负责一个互不相关的专业领域。Murray 负责评估函数和开本,Joe 负责主程序的所有其余部分,包括非常复杂的并行搜索代码,而我负责硬件和测试程序。团队的组建,没有一个人担任组长的角色,但所有人都是平等的。通常,每个人都专注于自己的专业领域。然而,我们有两次加入其中,帮助其中一位有需要的人。第一次是在第一个深蓝国际象棋芯片即将完成时。默里和乔出力帮助我。现在,随着乔尔不断发现的评估问题,以及新的评估职能的完成,穆雷已经无法独自承担这些工作量。我仍然把大部分时间花在新的国际象棋芯片上,所以乔提供了急需的帮助。穆雷和乔非常互补。

Up to this point in time, the Deep Blue team acted as a troika with each team member responsible for a disjointed area of expertise. Murray was responsible for the evaluation function and the opening book, Joe had all the rest of the main program, including the very complicated parallel search code, and I was responsible for the hardware and the test programs. The team was set up so that no one person had the role of the team leader but all were equal. Usually, everyone kept to his own area of expertise. On two occasions, however, we joined in to help one of us who was in need. The first time was when the first Deep Blue chess chip was near completion. Murray and Joe chipped in to help me. Now, with the constant stream of evaluation problems uncovered by Joel, and a new evaluation function to complete, Murray could not handle the workload all by himself. I was still spending most of my time on the new chess chip, so Joe provided the much-needed help. Murray and Joe complemented each other extremely well.

默里和乔的编程风格截然不同。默里喜欢在编写程序之前仔细考虑事情。这意味着他往往会取得稳定但较慢的进步。这一事实,加上我们的时间很短,就是为什么在 1996 年比赛时只使用了大约三分之一的硬件评估功能。乔更像是一位实验主义者。他会非常快地编写代码,尝试一下,看看它是如何工作的。随着所有新的评估功能突然可用,默里被财富淹没了。Joe 亲自负责新功能的评估函数代码,并很快在 Murray 已经编写的内容之上想出了新的东西。乔的国际象棋棋手比穆雷弱得多,但他在代码中放入的评估权重结果非常好。Murray 在评估函数方面的工作变得比编码更加微调。我们最终得到了乔的速度和穆雷的周到照顾的结合。

Murray and Joe had quite different programming styles. Murray liked to think things through before he wrote a program. This meant that he tended to make steady but slower progress. This fact, in conjunction with the short time that we had, was why only about one-third of the hardware evaluation features were in use at the time of the 1996 match. Joe was more an experimentalist. He would write code very fast, try it out and see how it worked. With all the new evaluation features suddenly available, Murray was overwhelmed with riches. Joe took it upon himself to work on the evaluation function codes for the new features, and quickly came up with new stuff on top of what Murray had already written. Joe was a much weaker chess player than Murray, but the evaluation weighting he put in the code turned out to be pretty good. Murray’s work on the evaluation function became more fine-tuning than coding. We ended up getting the combination of both Joe’s speed and Murray’s thoughtful care.

当新的国际象棋芯片从 VLSI Technology 回来时,我们已经有了相当好的评估功能。即使到复赛结束时,并非所有评估功能都充分发挥其潜力,但该功能已经比 1996 年比赛时的状态要好得多。穆雷使用新软件和新的国际象棋芯片与两个顶级商业 PC 程序进行了前几场比赛。当时,新软件仅使用一个国际象棋芯片,并以较慢的时钟速率运行。此外,由于幻影皇后问题,硬件搜索修剪被禁用,使新软件的速度又减慢了五到十倍。因此,“Pico”深蓝 Jr 的运行速度实际上只有每秒 100,000 到 300,000 个位置,或者与运行在高端 Pentium Pro 个人计算机上的最快商业国际象棋程序大致相同的搜索速度。这些程序之间最大的区别在于评估功能;那是,封装在程序中的国际象棋知识,或者就“Pico”深蓝 Jr 而言,封装在国际象棋硬件中。

When the new chess chips came back from VLSI Technology, we had a fairly good evaluation function in place. Not all the evaluation features were used to their full potential even up to the end of the rematch itself, but the function was already in much better shape than what we had at the time of the 1996 match. Murray played the first few games with the new software, using the new chess chip, against two of the top commercial PC programs. At the time, the new software used only one chess chip, running at a slowed down clock rate. Furthermore, because of the phantom queen problem, the hardware search pruning was disabled, slowing down the new software by another factor of five to ten. The “Pico” Deep Blue Jr was thus running at effectively only 100,000 to 300,000 positions per second, or roughly the same search speed as the fastest commercial chess program running on, say, a high end Pentium Pro-based personal computer. The biggest difference between the programs would be in the evaluation function; that is, the chess knowledge encapsulated in the programs or, in the case of “Pico” Deep Blue Jr, in the chess hardware.

穆雷在没有告诉任何人的情况下使用“Pico”深蓝Jr打了前几场比赛。有一天,他提到他曾与顶级商业国际象棋程序下过几盘棋,并让我猜猜结果。鉴于Pico在搜索速度上确实没有优势,而且新的评估函数还没有调好,我回答说:“大约50-50。” 随后穆雷抛出了重磅炸弹。皮科赢得了所有比赛。我想看看证据。穆雷向我展示了记忆中的比赛片段。皮科占据统治地位的原因显而易见。在我们寻求提高深蓝的国际象棋知识的过程中,我们创造了一些远远超出商业国际象棋程序的东西。在我看到的每个游戏片段中,商业程序的相对国际象棋知识缺陷变得一目了然,我可以准确地判断出每个游戏中发挥了哪些硬件评估功能。

Murray played the first few games with the “Pico” Deep Blue Jr without telling anybody. One day, he mentioned that he had played a few games against the top commercial chess programs, and asked me to guess the outcome. Given that Pico did not really have an edge in search speed, and the new evaluation function was still not tuned, I answered, “About 50–50.” Then Murray dropped the bombshell. Pico had won all the games. I wanted to see the evidence. Murray showed me fragments of the games from memory. The reason for the domination by Pico became obvious. In our quest to improve Deep Blue’s chess knowledge, we had created something way beyond what had been done in commercial chess programs. In each game fragment that I saw, the relative chess knowledge deficiency of the commercial programs became crystal clear, and I could tell precisely what hardware evaluation features were at play in each game.

此后穆雷又打了几场比赛。皮科继续获胜。在第十一场比赛中,穆雷决定为其中一个商业节目提供二比一的赔率。皮科输掉了比赛。另一个程序的相对战术优势对于 Pico 更好的评估功能来说已经无法弥补。此后不久,Joe 完成了使用多个国际象棋芯片的新 Deep Blue Jr 软件的第一个版本。我们决定是时候让新的深蓝小号去乔尔大师的优秀小象棋学校学习了。当Murray放弃与商业国际象棋程序的比赛时,Pico的10-0记录(不包括时间赔率的比赛)表明,在计算机对计算机比赛中,Pico比最好的至少强200分商业国际象棋程序,大约有百分之九十五的确定性。

Murray played a few more games afterwards. Pico kept on winning. On the eleventh game, Murray decided to give a two to one time odds to one of the commercial programs. Pico lost the game. The other program’s relative tactical strength became too much for Pico’s better evaluation function to compensate. Shortly thereafter, Joe completed the first version of the new Deep Blue Jr software that used multiple chess chips. We decided that it was the time for the new Deep Blue Jr to attend Grandmaster Joel’s Excellent Little Chess School. At the time when Murray abandoned the games against the commercial chess programs, the 10–0 record by Pico (excluding the game with time odds) indicated that, in computer vs. computer play, Pico was at least 200 rating points stronger than the best commercial chess programs, with roughly ninety-five percent certainty.

IBM 研究院传播部在 2 月初开始了复赛的媒体启动,大约在同一时间我们收到了新的国际象棋筹码。其中一项开球活动是二月下旬老深蓝Jr和新深蓝Jr之间的第一场比赛。

The Communication Department in IBM Research started the media kickoff for the rematch in early February, at roughly the same time we received the new chess chips. One of the kickoff events was the first game between the old Deep Blue Jr and the new Deep Blue Jr in late February.

目睹这场新旧之间一场非正式私人比赛的记者来自美国广播公司“夜线”节目和《时代》杂志。我们期待新版本获胜,但没有什么是理所当然的。毕竟,我们刚刚让新程序开始运行。比赛开局大致势均力敌,但随后新的 Deep Blue Jr 发起了一次看起来很有希望的攻击。老深蓝Jr并不关心,事实上对自己的机会相当乐观。新的深蓝Jr也变得乐观起来。两者都不可能是正确的。然后,突然间,新的深蓝Jr看到了材料的胜利。旧版本仍然认为位置大致相等,但绝对是错误的。它在新版本中发挥了意想不到的作用。现在,一切都结束了;新版本宣称它将赢得大量材料。新程序搬迁后,旧程序也看到了厄运。这场比赛只持续了大约二十步。所有人都惊叹不已。我不相信任何国际象棋实体可以对旧的深蓝Jr.做这样的事情。所有改进国王安全评估的芯片工作,以及所有利用硬件评估功能的新软件工作,都得到了丰厚的回报。《时代》杂志记者评价说,新深蓝Jr的招式具有明显的人类特征。加里将演奏一个与他 1996 年演奏的节目截然不同的节目。

The reporters witnessing the one-game informal private match, between the old and the new, were from the ABC ‘Nightline’ show and Time magazine. We were expecting the new version to win, but nothing could be taken for granted. After all, we only just had the new program working. The game started roughly equal, but then the new Deep Blue Jr initiated what looked like a promising attack. The old Deep Blue Jr was unconcerned, and was in fact quite optimistic about its own chances. The new Deep Blue Jr was also getting optimistic. Both could not be right. Then, suddenly, the new Deep Blue Jr saw the win of material. The old version still saw the position as about equal but was definitely wrong. It played a move unexpected by the new version. Now, it was all over; the new version declared that it would win massive material. After the move of the new program, the old program saw the doom as well. The game lasted only about twenty moves. Everyone was amazed. I would not have believed that any chess playing entity could do this to the old Deep Blue Jr. All the chip work on improving the king safety evaluation, and all the new software work to exploit the hardware evaluation features, had paid off handsomely. The reporter from Time commented that the moves played by the new Deep Blue Jr had a distinctly human-like quality to them. Garry would be playing a vastly different program from the one he played in 1996.

因此,新程序主导了其他计算机程序,包括旧版本的深蓝。这是一个好兆头,但仍有一个问题有待回答。新的深蓝对抗人类宗师会有多强?

So the new program was dominating other computer programs, including the old version of Deep Blue. It was a good sign, but one question remained to be answered. How strong would the new Deep Blue be against human Grandmasters?

米格尔·伊列斯卡斯大师是一位来自西班牙的国际象棋新星。在过去十年左右的时间里,西班牙已成为国际象棋大国。世界上最强的国际象棋锦标赛每年在西班牙利纳雷斯举行,米格尔是常客之一。由此,他与顶尖棋手也变得相当熟悉。我们第一次接触米格尔是在 1995 年,当时 Deep Thought II 输给了他两场比赛。我们邀请他帮助深蓝队为 1996 年的比赛做开场准备工作,他是在西班牙远程进行的。为了复赛,米格尔在1997年3月有两周的时间,他可以在这期间前往美国,我们抓住了这个机会并邀请了他过来。

Grandmaster Miguel Illescas was an up-and-coming chess star from Spain. In the last ten years or so Spain has become a major chess country. The strongest chess tournament in the world took place annually in Linares, Spain, with Miguel being one of the frequent participants. As a result, he became quite familiar with the top chess players. We had our first contact with Miguel in 1995 when Deep Thought II lost a two-game match to him. We invited him to help with Deep Blue’s opening preparations for the 1996 match, and he did it remotely from Spain. For the rematch, Miguel had a two-week time slot, in March 1997, in which he could travel to the United States, and we took the opportunity and invited him over.

米格尔花了两周的时间和乔尔一起准备开局,但偶尔他也会和小深蓝打几场休闲比赛,大部分比赛他都输了。这是另一个好兆头,但休闲游戏不能太认真。

Miguel spent the two weeks working with Joel on the opening preparation, but every once in a while he would play a few casual games with Deep Blue Jr. He lost most of the games. This was another good sign but casual games cannot be taken too seriously.

乔尔和米格尔合作得很好。乔尔对深蓝的行为方式了如指掌,而米格尔则处于国际象棋的顶尖梯队。当乔尔遇到一些困难的开局问题时,米格尔有时会想出顶级球员之间就同样的问题讨论过的内容。当米格尔对深蓝有疑问时,乔尔就会用大师们通用的国际象棋语言告诉他答案。

Joel and Miguel worked very well together. Joel had intimate knowledge of how Deep Blue behaved, while Miguel was in the top chess echelon. When Joel had some difficult opening problems, Miguel sometimes would come up with what the top players had discussed among themselves about the very same problems. When Miguel had questions about Deep Blue, Joel would tell him the answers in the common chess language among the Grandmasters.

米格尔来访后,大师尼克·德弗米安(Nick Dermian)和约翰·费多罗维奇(John Fedorowicz)受邀帮助深蓝在比赛中可能使用的特定开局系统。我们对这些会议严格保密,直到复赛开始,以避免向加里发出深蓝可能会打哪些空位的信号,从而给他带来危险的优势。准备一年多的时间,煞费苦心地改进计划,结果却被加里著名的主场开幕准备工作所屠杀,这是毫无意义的。我们参加比赛是为了对阵加里,而不是他的主场备战。和米格尔一样,尼克和约翰也和深蓝小号玩过休闲游戏,但大部分都输了。这又是一个好兆头。但可以肯定的是,我们需要在严肃的训练赛中与其他大师进行比赛。

After Miguel’s visit, Grandmasters Nick DeFirmian and John Fedorowicz were invited to help with specific opening systems that Deep Blue might use during the match. We kept these meetings a closely guarded secret until the start of the rematch to avoid signaling Garry which openings Deep Blue might play, giving him a dangerous edge. There was no point in preparing for over a year, painstakingly improving the program, only to get slaughtered by Garry’s renowned home opening preparations. We were going into the match to play Garry, not his home preparations. Just like Miguel, Nick and John also played casual games against Deep Blue Jr and lost most of them. Again, a good sign. But to be certain, we would need to play against other Grandmasters in serious training matches.

1996 年 4 月,我们为 Deep Blue Jr 安排了两场训练赛。第一场是对阵大师 Larry Christiansen,第二场对阵大师 Michael Rohde,两人都有丰富的计算机对战经验。

We scheduled two training matches for Deep Blue Jr in April 1996. The first was against Grandmaster Larry Christiansen, and the second against Grandmaster Michael Rohde both of whom had extensive experience playing against computers.

约翰和尼克都期待着亲自观看训练比赛。用约翰的话来说,“我希望看到他们受苦。” 拉里和迈克尔是他的亲密朋友。为了避免在复赛前他们的参与被泄露,他们决定在家工作。

John and Nick were both looking forward to seeing the training matches in person. In John’s words, “I would like to see them suffer.” Larry and Michael were close friends of his. To avoid any possible leaks of their involvement before the rematch, it was decided that they would work at home.

拉里参加了两场两场训练赛中的第一场。在第一局比赛中,深蓝Jr设置了一个很有希望的位置,然后突然放弃了一个棋子,没有明显的棋理原因。事实证明这是一个软件错误导致它随机走棋。ChipTest 第一次出局的阴影,其中 ChipTest 下了棋步生成器的第一步棋,无论好坏。Deep Blue Jr 错误与时间控制的处理方式有关。该错误是最近才引入的,乔已经尝试追踪它一段时间了。他以为它已经修好了,但显然它还在那里。这个局面看起来毫无希望,但不知何故,深蓝小号施展了一些战术魔法,结果是重复的平局。乔在第二天的比赛前找到了这个错误并修复了它。深蓝Jr赢得了第二局,并以1.5比0.5的比分赢得了比赛。与Michael的比赛没有出现任何bug,深蓝Jr再次以1.5比0.5的比分获胜。

Larry played the first of the two two-game training matches. In the first game, Deep Blue Jr set up a promising position, and then suddenly gave up a pawn for no apparent chess reason. It turned out to be a software bug causing it to play a random move. Shades of ChipTest’s first outing where ChipTest played the first move from the move generator, whether good or bad. The Deep Blue Jr bug was related to how time control was handled. The bug had been fairly recently introduced and Joe had been trying to track it down for a while. He thought that it was fixed, but apparently it was still there. The position looked hopeless, but somehow Deep Blue Jr whipped up some tactical magic and it was a repetition draw. Joe located the bug and fixed it before the game on the second day. Deep Blue Jr won the second game, and won the match by the score of 1.5 to 0.5. No bug showed up in the match against Michael, which again was won by Deep Blue Jr by the score of 1.5 to 0.5.

在常规时间控制下的比赛中,深蓝 Jr 对阵两位大师的表现都在 2700 多分左右,或者说与加里·卡斯帕罗夫 (Garry Kasparov) 的评级值相差 50 左右。游戏太少了,无法认真对待性能,但深蓝本身会比深蓝 Jr 快五到十倍,并且可能会强 100 到 200 个评级点。如果说深蓝还没有比加里强的话,那也已经是在攻击范围之内了。

The performance of Deep Blue Jr against the two Grandmasters in games played under regulation time control was in the high 2700s, or within about 50 rating points of Garry Kasparov. There were too few games to take the performance seriously, but Deep Blue itself would be five to ten times faster than Deep Blue Jr, and possibly 100 to 200 rating points stronger. If Deep Blue was not yet stronger than Garry, it was well within striking distance.

1997年决赛使用的深蓝2芯片。

The Deep Blue 2 chip used in the 1997 final match.

    1997 年版的《深蓝》会以不同的方式处理结局,并且有很大的获胜机会。

    The 1997 version of Deep Blue would have treated the ending differently, and had a decent chance to win.

    另一方面,比赛的最后两局确实表明仍然存在差距,而1996年的深蓝正如我们所怀疑的那样,并不真正值得获胜。回想起来,我很高兴我们输掉了费城的比赛。更好的玩家获胜。

    On the other hand, the last two games of the match did show that there was still a gap, and the 1996 Deep Blue, as we suspected, did not really deserve to win. Looking back, I am glad that we lost the Philadelphia match. The better player won.

    “印象”是广告中使用的术语。仅听到、看到或阅读广告信息并不能算作印象。只有当信息被潜在客户记住后,“印象”才算数。每次印象通常需要多次曝光该消息。

    “Impression” is a term used in advertising. Just hearing, seeing or reading the advertising message does not count as an impression. Only after the message is registered in the mind of the potential customer does the “impression” count. Each impression usually requires several exposures to the message.

 

 

第12章

CHAPTER 12

圣杯

The Holy Grail


盔甲中的中国佬

The Chinks in the Armor

在 1997 年重赛之前,我们在比赛前几个月就感受到了一种截然不同的气氛。我们对自己的机会感觉很好。加里·卡斯帕罗夫(Garry Kasparov)很强,但他并不是无敌的,深蓝正在逼近。我们正在快速而稳定地进步。我们自己在复赛前的评估是,如果深蓝不是已经比加里强的话,已经足够接近了,他赢不了超过一局。另一方面,他的实力足够强,我们严重怀疑我们能否以一局以上的优势赢得比赛。这场比赛也有很大的机会打平。从我们的角度来看,这将是一个非常好的结果。如果我们还有六个月的准备时间,如果我们能保持这样的进步速度,深蓝能够以一局以上的优势赢得比赛是可以想象的。但是,当然,我们没有六个月的时间了。无论我们想出什么,都必须要做,而且看起来新的深蓝可能已经足够好了。

Leading up to the 1997 rematch, we sensed a quite different atmosphere several months before the event itself. We felt good about our chances. Garry Kasparov was strong, but he was not invincible, and Deep Blue was closing in. We were making fast and steady progress. Our own assessment right before the rematch was that, if Deep Blue was not already stronger than Garry, it was close enough and he would not be able to win by more than one game. On the other hand, he was sufficiently strong that we seriously doubted that we could win the match by more than one game. There was a fairly good chance that the match would be drawn as well. From our point of view, this would be a perfectly good result. If we had, say, six more months to prepare, and if we could maintain our rate of improvement, it was conceivable that Deep Blue would be able to win the match by more than one game. But, of course, we did not have six more months. Whatever we came up with would have to do, and it looked as though the new Deep Blue might be good enough.

在公开场合,我们一直保持沉默,直到复赛前大约三个月。新的国际象棋筹码一直处于保密状态,由于我们不确定它们是否会发挥作用,因此直到比赛开始后它们的存在才被公开。我们制定了一个应急计划,如果我们在复赛前发现严重问题,就会恢复使用旧的国际象棋筹码。

Publicly, we kept quiet until about three months before the rematch. The new chess chips were kept under wraps, and since we were not absolutely sure that they would work out, their existence was not publicly known until the match was under way. We had a contingency plan to revert to the old chess chips if we discovered something terribly wrong before the rematch.

与此同时,加里显然变得紧张起来。1997年2月下旬,目睹了两个版本的深蓝Jr之间一场比赛的时代记者迈克尔·克兰茨(Michael Krantz)采访了他。迈克尔向加里提到他有观看了这款游戏,并对新程序所玩的类人动作质量印象深刻。加里变得非常激动。首先,他向迈克尔询问比赛得分。当他回答说他没有时,加里问他是否记得播放的开场曲。当加里发现迈克尔根本不知道这个开局叫什么时,他感到非常失望1。加里当时没有意识到,ABC Nightline有一盘完整的比赛录像带,而且事实上,早些时候已经播出了比赛的片段。加里知道ABC 夜线磁带吗?我不知道。但如果他这么做了,他就会知道自己确实有什么值得担心的。

Garry, meanwhile, was apparently getting nervous. In late February 1997, Michael Krantz, the same Time reporter who had witnessed the one-game match between the two versions of Deep Blue Jr, interviewed him. Michael mentioned to Garry that he had seen the game and been quite impressed with the human-like quality of the moves played by the new program. Garry became quite agitated. First, he asked Michael for the game score. When he replied that he did not have it, Garry asked whether he remembered what opening was played. Garry was quite disappointed when he found out that Michael did not have any idea what the opening was called1. What Garry did not realize at the time was that ABC Nightline had a complete video tape of the game, and in fact, had broadcast segments of the game earlier on air. Did Garry ever find out about the ABC Nightline tape? I don’t know. But if he did, he would have known that he indeed had something to worry about.

Garry 接受《时代》杂志采访后不久,我们收到 Frederic Friedel 发来的一封电子邮件,询问他是否可以与一位德国记者一起参观 IBM 研究中心。我们与弗雷德里克有着友好的关系,但我们担心他可能有调查与加里的比赛准备相关的信息的副业。我们会非常有礼貌但保持警惕,以避免给加里带来优势。Frederic于1997年3月上旬参观了该实验室,参观和采访都很顺利。在访问结束时,弗雷德里克发表了一个不同寻常的评论:“如果加里下反电脑象棋,他会轻松获胜。但他永远不会这么做。” 在反计算机国际象棋中,人类棋手故意采取次优的棋步,以达到他们认为计算机难以处理的局面。从这个意义上说,加里在1996年的比赛中就已经下过反电脑棋了。我们没有过多考虑弗雷德里克的评论。我们已经为与反计算机国际象棋比赛的可能性做好了充分的准备。我们不知道加里在即将到来的复赛中实际上会走多远。毫不奇怪,弗雷德里克的问题之一是我们是否认为深蓝能够赢得比赛。当我们直接回答他“深蓝很有可能获胜”时,他看起来不太高兴。我没有告诉他我相信深蓝至少有百分之五十以上的机会打平。

Shortly after Garry had his Time interview, we received an email from Frederic Friedel asking if he could visit IBM Research with a reporter from Germany. We had a friendly relationship with Frederic, but we were concerned he might have the side mission of probing for information relevant to Garry’s match preparation. We would be very polite but alert, to avoid giving Garry an edge. Frederic visited the lab in early March 1997, and both the visit, and the interview went fine. At the end of his visit, Frederic made an unusual comment, “Garry would win easily if he should play anticomputer chess. But he will never do it.” In anti-computer chess, the human player deliberately plays sub-optimal moves to get to positions that they think are difficult for the computer to handle. In this sense, Garry had already played anti-computer chess in the 1996 match. We did not give Frederic’s comment much thought. We had prepared extensively for the possibility of playing against anti-computer chess already. We had no idea how far Garry would actually go in the coming rematch. One of Frederic’s questions was, not surprisingly, whether we thought Deep Blue could win the match. He did not look too pleased when we gave him the straight answer that Deep Blue had a very real chance of winning. I did not tell him that I believed Deep Blue would have a better than fifty percent chance of at least drawing the match.

随着复赛的临近,加里开始通过他的经纪人提出各种要求,这些要求要么超出了已经签署的比赛合同,要么与已经签署的比赛合同相冲突。加里的一些要求相当合理,但另一些则不太合理。

As the rematch drew near, Garry started making various requests through his agents either outside of, or in conflict with, what was in the already signed match contract. Some of Garry’s requests were quite reasonable, but others were less so.

其中一个要求是使用一款新的国际象棋时钟,该时钟基于加里的规格,并由加里在广告中认可的瑞士钟表制造商爱彼定制设计。另一个与新的国际象棋时钟相关的是使用加里最近梦想的一种全新的、不规则的时间控制。比赛合同中已经明确规定要采用常规时间控制,就像1995年卡斯帕罗夫对阵阿南德世界冠军赛那样。最终,加里的新时钟被采用,但他的新时间控制却没有被采用。这种新的时间控制极其复杂,甚至连新的时钟都无法与之配合。更严重的是,即使加里的时钟能够处理他新的时间控制,深蓝团队也没有时间和精力修改程序来处理他迟来的请求。

One request was to use a new chess clock, based on Garry’s specifications and custom designed by Audemars Piguet, a Swiss watchmaker that Garry endorsed in advertisements. Another one, related to the new chess clock, was to use an entirely new and irregular time control that Garry had recently dreamt up. The match contract already specified that the regulation time control was to be used, just as in the 1995 Kasparov vs. Anand World Championship Match. In the end, Garry’s new clock was adopted, but not his new time control. This new time control was exceedingly complicated, and even the new clock did not work with it. More seriously, even if Garry’s clock could handle his new time control, the Deep Blue team had neither the time nor the energy to modify the program to deal with his late request.

加里自己并没有处理这个请求,所以他没有在这上面浪费时间和精力。但如果我们同意了,我们将不得不花费宝贵的时间(我们实际上并没有这样的时间)来编写新的时间控制并进行大量的测试,这可能会花费我们数周的准备时间。请记住,我们只有不到 12 周的时间来为复赛准备新的深蓝。出于善意,我们确实花了时间查看新的国际象棋时钟和新的时间控制文件。团队认为我们无法在新的时间控制下编程并及时完成比赛准备。然而,IBM 研究中心的乔治·保罗被要求评估新的国际象棋时钟,看看是否有任何方法可以满足加里的两个要求。在花了几天时间检查时钟后,当他报告说它甚至在处理常规时间控制时也存在足够的问题时,我们松了口气。乔治汇报后,爱彼及时为复赛定好时钟,进行常规时间控制。

Garry was not handling the request himself, so he was not wasting his time or energy on it. But if we had agreed to it, we would have had to spend valuable time, which we did not really have, to code the new time control and do extensive testing which could have cost us weeks of preparation time. Remember that we had less than 12 weeks to bring up the new Deep Blue for the rematch. We did, out of good will, spend time looking over the new chess clock and the document for the new time control. It was the team’s opinion that there was no way we could program in the new time control and finish the match preparation in time. However, George Paul from IBM Research was asked to evaluate the new chess clock, and see whether there was any way that both of Garry’s requests could be satisfied. We were relieved when, after spending days examining the clock, he reported that it had enough problems handling even the regular time control. After George’s report, Audemars Piguet fixed the clock in time for the rematch, with a regular time control.

在复赛合同谈判初期,双方同意,如果加里提出平局,运营商将被允许将此提议告知深蓝。如果球队决定拒绝抽签提议,深蓝就会被告知抽签提议已被拒绝。这条规则的目的是为了防止重演1996年比赛中的情况,当时Mike Valvo在深蓝出招后,在不知道已经商议平局的情况下,迫使深蓝队做出了草率的决定。就我们而言,能够超越深蓝并接受抽奖邀请的能力是出于对人类玩家的尊重的礼貌。国际象棋有太多的例外,不可能对计算机进行编程,使其了解在每种情况下都会绘制一个位置。如果程序员没有能力推翻,人类对手可能会被迫不必要地玩出既定的结局,这除了让他们疲惫不堪之外没有任何用处。随着复赛的临近,加里有了新的想法,在同意新规则之前,想再加一条规则,即深蓝战队在审议抽签方案时,不能看到深蓝队是如何评估排名的。这要求太多了,因为我们的审议将在没有深蓝玩家的任何输入的情况下进行。毕竟这场比赛应该是在加里和深蓝之间进行的。如果深蓝取得了人类无法看到的辉煌胜利怎么办?抽签的匹配规则从来没有得到让任何人满意的解决。在内部,深蓝团队做出了一个简单的决定:我们不会接受任何位置不明确的抽签报价。复赛时,他没有在位置不明确的地方打平局,所以这个问题就没有出现。

Early in the contract negotiation for the rematch, it was agreed that, if Garry offered a draw, the operator would be allowed to tell Deep Blue of the offer. If the team decided to decline the draw offer, Deep Blue would then be told that the draw offer had been declined. The purpose of this rule was to prevent the repeat of the situation in the 1996 match where Mike Valvo forced the Deep Blue team to make a rash decision after Deep Blue had come up with a move, without knowing that a draw offer was being deliberated. As far as we were concerned, the ability to be able to override Deep Blue, and accept draw offers, was a courtesy out of respect for the human player. Chess has too many exceptions, and it is not possible to program the computer to understand that a position is drawn in each and every circumstance. Without the ability of the programmers to override, the human opponents might be forced to needlessly play out drawn endings, which would serve no useful purpose other than tiring them. As the rematch drew near, Garry had second thoughts, and before agreeing to the new rule, wanted to add another rule that the Deep Blue team would not be allowed to see how Deep Blue assessed the position while deliberating on the draw offer. This was asking for too much as our deliberation would then be made without any input from Deep Blue the player. The match was supposed to be between Garry and Deep Blue, after all. What if Deep Blue finds a brilliant win that no human being would see? The match rule for draw offers was never resolved to anyone’s satisfaction. Internally, the Deep Blue team made the simple decision that we would not accept any draw offers in unclear positions. In the rematch, he did not offer draws in unclear positions, so the problem never came up.

自 1986 年我们推出第一款 ACM 游戏以来,ChipTest 及其后继者已经公开玩了一百多场正式锦标赛或比赛,而 Garry 已经或本来可以参加其中的每一场比赛。比赛合同明确要求深蓝团队应要求提供我们所有公开锦标赛比赛的分数。这些分数很容易获得,因为我们处于计算机国际象棋食物链的顶端,并且我们的游戏总是受到仔细审查。但是,在我的记忆中,加里从来没有费心去询问比赛的比分。然而,这并没有阻止他一再抱怨自己无法玩任何深蓝游戏。确实,他无法访问该程序最新版本所玩的游戏,但通常他也无法访问新的和改进的阿南德在世界冠军赛前几个月的训练期间所玩的游戏。有人可能会说,该程序的新版本与旧版本有很大不同,因此加里应该有机会研究它。这是一种倒退的说法。新版本之所以有很大不同,只是因为我们投入了大量的工作,就像阿南德和加里本人所做的那样。阿南德是否应该仅仅因为加里在六个月的训练中取得了进步就被迫给他提供样本比赛?如果他要求阿南德给他一些游戏样本来研究,任何有理智的人都会认为这个要求很荒唐。为什么他是唯一拥有比赛秘密武器的人?

Since our first ACM game in 1986, ChipTest and its successors had played over one-hundred formal tournament or match games in public, and Garry had or could have had access to every one of them. The match contract explicitly obliged the Deep Blue team to provide, when requested, the scores of all our public tournament games. These scores were readily available, as we were at the top of the computer chess food chain, and our games were always carefully scrutinized. But, in my recollection, Garry never bothered to request the game scores. This, however, did not stop him from repeatedly complaining that he had no access to any Deep Blue games. True, he did not have access to games played by the newest version of the program, but usually neither would he have had access to games played by, say, the new and improved Anand during the months of training before their world title match. One might argue that the newer version of the program was vastly different from the old one, so Garry should have been given a chance to study it. This is a backward argument. The newer version was vastly different only because we had put a tremendous amount of work into it, just as Anand and Garry himself would have done. Should Anand be forced to give Garry sample games just because he had improved himself from six months of training? If he had asked Anand to give him sample games to study, any reasonable person would have considered the request ludicrous. Why should he be the only one to have secret weapons for the match?

我们想要公平,但在赛前训练中给加里比赛会让我们处于严重劣势。在 1989 年他与 Deep Thought 的表演赛中,他施展了毁灭性的新开局,在第二场比赛中打破了 Deep Thought 的开局。当《深思》一到达时就已经死了,因为它发现自己已经脱离了它的开篇书。如果加里知道深蓝会玩什么开局,谁能说历史不会重演呢?我们想对阵世界冠军,而不是他在我们开局时的主场比赛。尽管如此,我们能否给他一些象征性的游戏分数?理论上是的。但会有一些严重的并发症。如果他不能利用比赛比分,那么我们就会被指控故意误导他。如果他能够利用比赛分数,那么,为了让比赛变得公平,我们就必须对开局书进行大幅修改,改变程序,甚至可能建造一台新机器(这样才能很好地发挥新的空缺) 2 . 然后加里会要求新游戏。回到原点。

We wanted to be fair, but to give Garry games from our pre-match training would have placed us at a serious disadvantage. In the exhibition match between him and Deep Thought in 1989, he had unleashed a devastating new opening line that busted Deep Thought’s opening in game two. Deep Thought was dead on arrival when it found itself out of its opening book. If Garry had some idea of what openings Deep Blue would play, who is to say that history would not be repeated? We wanted to play the World Champion, not his home preparation against our openings. Could we have given him some token game scores nonetheless? Theoretically, yes. But there would have been some serious complications. If he had not been able to take advantage of the game scores, then we would have been accused of deliberately misleading him. If he had been able to make use of the game scores, then, in order to make the match a fair one, we would have had to make substantial changes to the opening books, change the program, and possibly even build a new machine (so that it could play new openings well)2. Garry would then ask for new games. Back to square one.

加里在 1996 年的比赛中也曾发出过类似的抱怨,但强度并不完全相同。为什么他现在已经和深思交了两次,深蓝交了六次了,现在还抱怨自己没有深蓝的玩法可以研究呢?一个可能的原因是,在他看来,他应该有新的游戏来研究,尽管这不是他对抗人类对手的权利。还有另一种可能性。Frederic Friedel 在访问 IBM 研究院期间发表了一段有趣的言论,或许能提供一些启示。在1997年重赛之前,加里在顶级国际象棋锦标赛中与其他顶级人类棋手取得了巨大成功,似乎正处于职业生涯的巅峰。据弗雷德里克说,这一巨大成功的秘密之一是他在商业国际象棋程序的帮助下改进了他的整个开局曲目。这些程序被用来指出他的棋谱中潜在的问题,并找到新的动作来让对手大吃一惊。深蓝是一个比任何程序都强大得多的国际象棋机器加里有权访问。如果他能在开局准备中使用商业国际象棋程序取得很好的效果,那么在深蓝的帮助下,一群大师的合作肯定会产生一些巨大的开局惊喜。加里,这位有史以来最伟大的国际象棋开局理论家之一,真的害怕我们家里的开局准备吗?

Garry had made similar complaints during the 1996 match, but they were not quite of the same intensity. Why did he still complain now, when he had already played against Deep Thought twice and Deep Blue six times, that he did not have any Deep Blue game to study? One possible reason was that, in his mind, he deserved to have new games to study, even though it was not a right that he would have against a human opponent. There was another possibility. During Frederic Friedel’s visit to IBM Research, he made an interesting remark that might shed some light. Before the 1997 rematch, Garry had had great success in top-level chess tournaments against other top human players, and appeared to be at the peak of his career. One of the secrets of this great success, according to Frederic, was that he was revamping his entire opening repertoire with the help of commercial chess programs. These programs were used to point out potential problems in his repertoire, and to find new moves to surprise his opponents. Deep Blue was a far more powerful chess machine than any program that Garry had access to. If he could use commercial chess programs in his opening preparations to great effect, a group of Grandmasters working together with the help of Deep Blue surely would produce some great opening surprises. Was Garry, one of the greatest chess opening theorists ever, actually afraid of our home opening preparations?

加里在复赛中遇到的问题不仅仅是一般的紧张。对于一个对自己的机会感到紧张的人来说,令人惊讶的是,他严重低估了深蓝的实力。在赛前的新闻发布会上,有记者问他,深蓝相对于商业国际象棋程序的实力如何。他回答说,他希望深蓝能够在与最好的商业国际象棋程序的十场比赛中赢得八场。根据我们掌握的实验数据,“Pico”深蓝Jr比深蓝慢了大约一千倍,在与最好的商业程序的十场比赛中,它可能会赢得至少八场。两个相同程序之间十倍的速度差异很可能会产生 8 比 2 的胜率,我们正在谈论一个已经能够赢得十场比赛中的八场的基础程序,加速了一千倍……当我听到加里的声明,我对自己说,“孩子,卡斯帕罗夫一定会感到震惊。”

Garry’s problems going into the rematch went beyond just general nervousness. For someone nervous about his own chances, he, surprisingly, grossly underestimated Deep Blue’s strength. In the pre-match press conference, a reporter asked him how strong he thought Deep Blue was relative to the commercial chess programs. He answered that he expected Deep Blue to win eight out of ten games against the best commercial chess programs. “Pico” Deep Blue Jr, which was about one-thousand times slower than Deep Blue, probably would win at least eight out of ten games against the best commercial programs, based on the experimental data we had. A ten-fold difference in speed between two identical programs would very likely produce a winning ratio of 8 to 2, and we are talking about a base program that was already capable of winning eight out of ten games, speeded up a thousand times … When I heard Garry’s statement, I said to myself, “Boy, Kasparov is in for some shock.”

1996年比赛之前,加里看上去是不可碰触的,而在1996年比赛结束时,他也是不可碰触的。然而,在 1997 年复赛之前,中国佬就已经出现在他的盔甲里了。我喜欢我们的机会。

Before the 1996 match, Garry looked untouchable, and at the end of the 1996 match, he was untouchable. Before the 1997 rematch, however, chinks were already showing up in his armor. I liked our chances.

开设商店

Setting Up Shop

复赛在纽约市曼哈顿岛举行,距离我们实验室大约半小时车程。理论上,我们可以从家里通勤,但我们大多数人在整场比赛中都呆在曼哈顿。乔尔住在曼哈顿下城,但和我们其他人一起住在米开朗基罗酒店,该酒店就在曼哈顿上城比赛场地公平大厦的街对面。赛前从西班牙赶来的米格尔也和妻子一起入住了米开朗基罗酒店。米格尔、尼克和约翰在新闻发布室担任记者的特级评论员。当乔尔在比赛期间接受记者采访时,米格尔也是我们的副手。尼克和约翰是土生土长的纽约人,职责较少,每天晚上都回家。所有的大师参与了比赛之间的开场准备工作。IBM 在米开朗基罗酒店设立了一个客户接待区,并提供来自比赛现场的微波视频。受邀的IBM客户可以在舒适轻松的环境中观看比赛和现场解说。食物和酒不断供应。尼克和约翰都对美酒有着浓厚的兴趣,经常在顾客接待区看到他们的身影。赛后,加里声称他们是为米开朗基罗的深蓝队效力。尼克和约翰只是在那里享受美食和美酒,享受他们的美食乐趣,并与 IBM 客户交往。

The rematch took place on the Manhattan Island of New York City, roughly half an hour drive from our lab. In theory, we could have commuted from our homes, but most of us stayed in Manhattan throughout the entire match. Joel lived in lower Manhattan, but stayed with the rest of us in the Michelangelo Hotel, which was just across the street from the Equitable Building, the match site, in upper Manhattan. Miguel, who arrived from Spain just before the match, also stayed in the Michelangelo with his wife. Miguel, Nick, and John served as Grandmaster commentators for the reporters in the pressroom. Miguel was also serving as our second when Joel had to give interviews to reporters during the games. Nick and John, having fewer responsibilities and being native New Yorkers, went home every night. All the Grandmasters were involved in the opening preparation sessions between games. IBM set up a customer hosting area in the Michelangelo Hotel, with a microwave video feed from the match site. The invited IBM customers could watch the games and the live commentaries in a comfortable and relaxed setting. Food and wine were served continuously. Nick and John, both with a taste for fine wine, were frequently sighted at the customer hosting area. After the match, Garry made the outrageous claim that they had played for Deep Blue from the Michelangelo. Nick and John were just there to enjoy the food and the wine, taking care of their gastronomic delight, and rubbing shoulders with the IBM customers.

在纽约市举办复赛的决定一开始就造成了一些难题。比赛谈判早在一年多前就开始了,但要找到规模合适、设施齐全、空置的场地已经相当困难。起初,千禧酒店被宣布为比赛场地,但在复赛宣布后,酒店表示已预订复赛最后一个周末举行婚礼,并希望IBM支付额外费用,以便重新安排婚礼。恰巧IBM首席执行官郭士纳经常在Equitable Building召开股票分析师会议,不少IBM人都熟悉这个地方。Equitable 大楼是 Equitable Insurance 的总部,是一栋现代化的建筑,楼层之间的视频连接所需的所有布线均已就位。经过广泛检查后,认为新址是更好的场地,千禧酒店的预订被取消。不知道复赛开始后,千禧酒店的管理人员看到如此大的宣传有何感想?

The decision to host the rematch in New York City created some difficult problems early on. The match negotiation started more than a year beforehand, but it was already quite difficult to find a venue that was of the right size, had adequate facilities, and proper vacant time slot. At first, the Millennium Hotel was declared as the match site but after the rematch announcement, the hotel said that a wedding ceremony was booked for the last weekend of the rematch, and wanted IBM to pay an additional fee so that the ceremony could be rescheduled. It so happened that IBM CEO Lou Gerstner usually hosted his stock analyst meetings at the Equitable Building, and quite a number of IBMers were familiar with the site. The Equitable Building, the headquarters of Equitable Insurance, is a modern building with all the necessary wiring for video connections between floors already in place. After an extensive examination, the new site was decided to be the better venue, and the Millennium Hotel booking was cancelled. I wonder how the managers of the Millennium Hotel felt after seeing the massive publicity, once the rematch got under way?

公平大厦是一座雄伟的摩天大楼,很适合公平保险公司这样的大公司的总部。百老汇剧院区就在附近,在休息日期间,为制作非百老汇国际象棋表演的 IBM 团队获得了演出门票。不过,我们深蓝团队中没有人去看演出。乔、默里和我在这个项目上总共工作了近三十年,我们打算密切关注这一问题。百老汇演出可能要等一段时间。

The Equitable Building was an imposing skyscraper, as befits the headquarters of a big corporation like Equitable Insurance. The Broadway Theater district was close by, and during one of the rest days, show tickets were secured for the IBM team producing the off-Broadway chess show. None of us on the Deep Blue team went to the show, though. Joe, Murray, and I had worked for nearly three decades, combined, on the project, and we intended to keep our eyes on the ball. Broadway shows could wait a while.

步行即可到达洛克菲勒中心。第五场比赛前的周六早上,一些有进取心的 IBM 员工在 Matt Thoennes 的带领下,将通常放在比赛礼堂舞台上的“真人大小”深蓝模型推到了街上就在洛克菲勒中心NBC “今日”节目现场直播的旁边。《今日报》的主持人非常惊讶地发现窗外有一个大黑匣子。

Within walking distance was the Rockefeller Center. The Saturday morning before game five, some enterprising IBMers, led by Matt Thoennes, pushed the “life-size” Deep Blue model, which normally was on the stage at the match auditorium, into the street right next to the live broadcast of the NBC’s ‘Today’ show in the Rockefeller Center. The ‘Today’ hosts were quite surprised to find the big black box outside their window.

公平大厦的主入口位于第七大道。穿过正门,观众们来到了一座三层楼高、气势磅礴的大理石大厅,阳光透过玻璃天花板照射进来。对面的墙上挂着一幅罗伊·利希滕斯坦的巨幅画作。由于俄克拉荷马城和世贸中心爆炸事件后的安全问题,观众被要求通过金属探测器和安全检查。然后,他们必须乘坐一架飞机前往拥有 480 个座位的礼堂,这里是解说室。并非所有座位都向公众开放。IBM、ACM 和 Garry 都有一些保留票。IBM 门票的一部分送给了美国国际象棋联合会的客人,其余大部分则送给了幸运的学童。少量 IBM 门票是为想要亲眼目睹礼堂的受邀客户保留的。在未超出消防安全容量的情况下,记者被允许进入礼堂。

The main entrance for the Equitable Building was on seventh Avenue. After passing through the main entrance, the spectators arrived in a majestic marble lobby, three stories high, with daylight coming in through the glass ceiling. Directly facing them on the far wall was a giant painting by Roy Lichtenstein. As a result of security concerns following the Oklahoma City and the World Trade Center bombings, the audience was asked to pass through metal detectors and a security check. They then had to go one flight down to the 480-seat auditorium, which served as the commentary room. Not all the seats were available to the general public. IBM, ACM, and Garry all had some reserve tickets. Some of the IBM tickets were given to the United States Chess Federation for their guests, and most of the others went to lucky school children. A small number of IBM tickets were reserved for invited customers who wanted to see the auditorium first hand. Reporters were allowed into the auditorium when the fire safety capacity was not exceeded.

观众席前部的三块巨型投影屏幕上实时显示棋盘、游戏室以及解说员使用的电脑分析板的视频直播。投影设备由 AmPro 免费提供,该公司在费城提供相同的服务。此前曾与松下幸之助进行过会谈。如果会谈成功,将使用视频墙代替投影屏幕,现场比赛视频将在时代广场的巨大松下/NBC 电视屏幕上播放。但松下发生了一些人事变动,谈判破裂。

At the front of the auditorium, three giant projection screens displayed live video feeds of the chessboard, the game room, and the computer analysis board used by the commentators. The projection equipment was provided gratis by AmPro, the company who provided the same service in Philadelphia. There had been talks earlier with Matsushita. Had the talks worked out, video walls would have been used instead of projection screens, and live match video would have been shown on the giant Panasonic/NBC TV screen in Time Square. However, some personnel changes took place at Matsushita, and the talks broke down.

舞台右侧的投影屏幕下方,矗立着一RS/6000 SP 超级计算机的“真人大小”模型。框架是一个冰箱大小的盒子,高超过六英尺。加里本人也可以很好地融入其中。真实的框架最多可容纳 16 个 RS/6000 工作站。舞台上的框架实际上只是真机的木箱模型。比赛结束后,这个木质模型被包机运往加利福尼亚州进行电视广告拍摄。

Below the projection screens, on the right side of the stage, stood a “life-size” model of one frame of an RS/6000 SP supercomputer. The frame was a refrigerator-sized box well over six feet high. Garry himself could have fitted nicely into it. A real frame would house up to 16 RS/6000 workstations. The frame on the stage was actually just a wooden box model of the real machine. After the match, this wood model was shipped to California on a charter plane for TV commercial shoots.

1996年那场比赛的裁判迈克·瓦尔沃(Mike Valvo)担任观众席的三位主要评论员之一。1996年比赛的主要评论员,GM Yasser Seirawan另外两个是 IM Maurice Ashley。在这三位评论员中,我认为亚西尔绝对是卡斯帕罗夫的支持者,而莫里斯和迈克则相对中立。1974 年罗伯特·伯恩(美国)输给鲍里斯·斯帕斯基(苏联)后,亚西尔是第一位成为世界锦标赛候选人的美国选手。如果我对这种偏见的看法是正确的,那么这有点有趣,特别是考虑到加里拒绝分享1997年的比赛中与他同台。亚西尔和加里在 20 世纪 80 年代末曾是亲密的朋友,但国际象棋政治阻碍了友谊。莫里斯第一次出名是在他执教的哈莱姆区初中国际象棋队 Raging Rooks 于 1991 年赢得美国初中国际象棋锦标赛冠军时。我第一次见到他是在他的团队赢得冠军后不久访问 IBM 研究中心时。除了他的教练活动之外,他给我留下深刻印象的一件事是他明确表示希望成为第一位非洲裔特级大师。复赛后不久,他的梦想就变成了现实。好人总会遇到好事。

Mike Valvo, the match arbiter for the 1996 match, was serving as one of the three main commentators in the auditorium. The main commentators from the 1996 match, GM Yasser Seirawan and IM Maurice Ashley, were the other two. Of the three commentators, I thought Yasser was decidedly pro-Kasparov, while both Maurice and Mike were relatively neutral. Yasser was the first American player to become a World Championship Candidate after Robert Byrne (USA) lost to Boris Spassky (USSR) in 1974. If I am right about this bias, it was kind of interesting, especially considering that Garry refused to share the same stage with him during the 1997 match. Yasser and Garry used to be close friends in the late 1980s, but chess politics got in the way of friendship. Maurice first became well known when Raging Rooks, a Harlem junior high school chess team he coached, won the US Junior High School Chess Championship in 1991. The first time I met him was when his team visited IBM Research shortly after winning the title. Besides his coaching activities, one thing that really impressed me about him was his stated desire to become the first Grandmaster of African descent. Not long after the rematch, his dream became reality. Good things do happen to good people.


比赛举办地公平大楼的三十五层禁止观众进入。那里只有一部普通电梯,保安只允许列入限制名单的人进入。乘电梯后,另一位保安会迎接你,他会根据名单检查你的名字。由于进去的人很少,几天后保安就会认出你的脸并微笑着跟你打招呼。

The thirty-fifth floor of the Equitable Building, where the games took place, was off limit to the spectators. Only one regular elevator goes up there, and only people on a restricted list were allowed by the security guard to enter it. After taking the elevator, you would be greeted by another security guard, who would check your name against the list. Since very few people were allowed, after a few days the security guard would recognize your face and greet you with a smile.

为了到达游戏室,嘉宾和记者都穿过一个储藏室,储藏室通过一条蜿蜒的临时走廊与游戏室相连。走廊上用隔音材料建造,隔绝了储藏室空调的噪音。游戏室有很多大功率灯具和视频设备,产生大量热量。储藏室空调为游戏室提供急需的空调,并且必须连续运行。进入游戏室,你会看到一排椅子。其中一把椅子华丽而舒适,看起来很吸引人。这是克拉拉亲自挑选的椅子。比赛期间 Clara 总是在场,IBM 很简单地礼貌地确保她尽可能舒适。除了椅子之外,你可以看到一个分隔区域,大约占据了游戏室的四分之一。这是加里的休息区,配有沙发、桌子、水果和瓶装水水。他喜欢吃香蕉,每天都会有新鲜、无瑕疵的香蕉送到餐桌上。同层还为他准备了一间隔音房,里面塞满了更多好东西。这次他的房间里没有厕所,可能是因为很难带一个厕所。他和三十五层的其他船员共用一个休息室,只不过他的优先权最高。 。当他在休息室时,任何人,包括深蓝团队的成员,都不得使用休息室。

To get to the game room, guests and reporters alike passed through a storage room that was connected to the game room by a winding temporary corridor. The corridor was constructed with sound proofing material to cut off the noise from the storage room air conditioner. The game room had many high-power lamps and video equipment that generated lots of heat. The storage room air conditioner provided much-needed air conditioning for the game room and had to run continuously. On entering the game room, you would see a single row of chairs. One of the chairs—ornate yet comfortable—looked quite inviting. This was Clara’s chair that she had personally chosen. Clara was always present during the games, and it was simple courtesy for IBM to make sure that she was as comfortable as possible. Beyond the chairs, you could see a partitioned area that occupied roughly a quarter of the game room. This was Garry’s rest area, furnished with a sofa, table, fruits and bottled water. He liked bananas, and every day, fresh, unblemished bananas would be delivered to the table. Another soundproof room was provided for him on the same floor, stuffed with more goodies. There was no port-a-john in his room this time, possibly because of the difficulty of bringing one in. He shared the same rest room as the rest of the crew on the thirty-fifth floor, except that he had the highest priority. Nobody, including members of the Deep Blue team, was allowed to use the rest room while he was in it.

游戏室里加里休息区的另一边是另一扇门,通常是关着的。这实际上是通往声音控制室的两扇隔音门中的第一扇。这也是加里的私人入口。两扇门之外是一条铺着窗帘的走廊。每当他打开第一扇双门时,视频摄制组和 IBM 团队就会被告知不要靠近。挂着窗帘的走廊通向加里的私人房间和公共休息室。从深蓝手术室到公共休息室的唯一通道也是通过帘子走廊。由于他再次拥有最高优先级,这意味着 IBM 团队中的任何人都不允许在他离开游戏室时去休息室,即使他没有使用游戏室。

Beyond Garry’s resting area in the game room was another door, usually closed. It was actually the first of two soundproofed doors leading to the sound control room. It was also Garry’s private entrance. Beyond the two doors was a draped corridor. Whenever he opened the first of the double doors, the video crew and the IBM team were informed to stay away. The draped corridor led to both Garry’s private room and the common rest room. The only way from the Deep Blue operational room to the common rest room was also through the draped corridor. As he again had highest priority, this meant that no one on the IBM team was allowed to go to the rest room whenever he was out of the game room, even when he was not using it.

游戏室的前半部分是游戏区。与1996年那场比赛的无菌环境相比,它看起来几乎是家常便饭。这个房间实际上是一个电视演播室,供公平保险公司的高级管理人员用来广播和录制公司信息,并且被设置成看起来像贵族或富有的律师的书房。前墙中央贴着一张复赛的大海报。如果海报后面有一幅画,那很可能是中世纪的场景,也许是一位骑士拯救陷入困境的少女,也可能是一幅全副武装的家族族长的肖像。墙壁上装饰着几幅中世纪场景的画作。书架上摆满了旧法律书籍。架子上有一个船模型,还有几个木鸭诱饵。盆栽植物让光线充足的房间充满生机。

The front half of the game room was the playing area. Compared to the sterile environment of the 1996 match, it looked almost homely. The room was really a TV studio used by senior executives of Equitable Insurance for broadcasting and recording corporate messages, and was set up to look like the study of an aristocrat or perhaps a rich lawyer. In the center of the front wall was a big poster for the rematch. If there was a painting behind the poster, it probably had a medieval scene, perhaps of a knight rescuing a damsel in distress, or maybe a portrait of the patriarch of the family in full battle regalia. Several paintings of medieval scenes adorned the walls. The bookshelves were full of old law books. On the shelves was a ship model as well as several wooden duck decoys. Potted plants livened up the well-lit room.

灯光对加里来说有点太亮了,但他妥协了,因为他想在视频中看起来很好。游戏桌上方的灯光必须调整好几次,直到他最终满意皇后周围不再有阴影为止。游戏桌是定制的,后部有一个四十五度的弯曲,朝向右侧的深蓝操作员。桌子的这个部分放置着一台价值 5000 美元的黑色 IBM 平板显示器。左边,加里坐在一张镶满铆钉的皮革扶手椅上,椅子的硬度正是他喜欢的。右边,深蓝操作员坐在一张类似的椅子上。比赛仲裁员卡罗尔·贾雷茨基 (Carol Jarecki) 坐在右侧的桌子旁。默里、乔和我在一场又一场的比赛中轮流担任操作员,在长时间的比赛中偶尔会换人。比赛规则禁止操作员在加里思考时离开椅子,因此操作员在球台上呆了很长时间后通常会感到相当不舒服。

The lighting was a little too bright for Garry but he compromised since he wanted to look good on the video. The lighting above the playing table had to be adjusted several times until he was finally satisfied that there was no more shadow around his queen. The playing table was custom made, with a forty-five-degree bend at the rear toward the Deep Blue operator on the right. A black, $5000 IBM flat panel display sat on this part of the table. To the left, Garry sat on a studded leather armchair of the exact firmness he liked. On the right, the Deep Blue operator sat on a similar chair. The match arbiter, Carol Jarecki, sat at a desk to the right. Murray, Joe and I rotated as the operator from game to game, with occasional substitution during long games. The match rules forbade the operator from leaving the chair while Garry was thinking, so the operator was usually fairly uncomfortable after staying at the table a long while.

深蓝本身也在三十五层。复赛前几个月,IBM RS/6000 部门宣布推出新的 RS/6000 SP,其 CPU(中央处理单元)的速度大约是我们在 1996 年比赛中使用的型号的 CPU 的两倍。对新产品的需求超过了供应。我们能够访问的唯一方法是获得在运送给客户之前在工厂车间进行测试的产品。与往常不同的是,所讨论的 RS/6000 SP 将在 Equitable 大楼的 35 层进行地面测试,在全世界的眼前。

Deep Blue itself was also on the thirty-fifth floor. A few months before the rematch, IBM RS/6000 division announced the availability of a new RS/6000 SP, which had CPUs (Central Processing Units) that were about twice as fast as the CPUs in the model we used for the 1996 match. The demand for the new one was outstripping the supply. The only way that we could have access was to get one that was being tested on the factory floor prior to being shipped to the customer. Instead of the usual, the RS/6000 SP in question would be doing the floor testing on the thirty-fifth floor of the Equitable Building, in front of the eyes of the world.

RS/6000 SP 在复赛前一周抵达公平大楼,所有棋牌都已就位。它有两个框架,每个框架容纳 15 个 RS/6000 节点,总共 30 个节点,即 30 个 RS/6000 工作站。每个SP节点包含两张棋牌,每张棋牌有八个筹码。系统中棋子总数为480个。理论最大搜索速度约为每秒10亿个位置。这是系统保证不超过的速度。实际最高速度约为每秒 2 亿个位置。搬迁前,负责物流的人员告诉我们,深蓝RS/6000 SP不会有任何外部电话访问。因此,在机器搬走的第二天,我们就搬进了米开朗基罗酒店,因为我们无法从约克镇的办公室或家中访问机器。

The RS/6000 SP arrived at the Equitable Building a week before the rematch, with all the chess cards in place. It had two frames, each frame housing 15 RS/6000 nodes for a total of 30 nodes, or 30 RS/6000 workstations. Each SP node contained two chess cards, with eight chips per card. The total number of chess chips in the system was 480. Theoretical maximum search speed was about one billion positions per second. This is the speed guaranteed-not-to-be-exceeded by the system. The actual maximum speed was around 200 million positions per second. Before the move, the people handling the logistics told us that there would not be any external phone access to the Deep Blue RS/6000 SP. So we all moved into the Michelangelo Hotel the day after the machine was moved, since there was no way for us to access the machine from our offices in Yorktown, or our homes.

当我们登上公平大厦的三十五层时,我们发现深蓝坐落在大楼远角一个壁橱大小的小空间里。机器有专用的空调,但周围的空间却相当温暖。相邻的空间稍微凉爽一些,几天后我们安装了落地扇来循环空气以帮助降低温度。即使对于穿着短袖T恤和短裤的人来说,整个机器区域也太热了。我们无法访问当时三十五层的其他部分,在深蓝上长时间工作是相当困难的,因为我们被迫呆在机器区域。几天来,当乔没有对并行程序的搜索效率进行自己的测试时,我对系统进行了广泛的测试。一些棋牌似乎在运输过程中受损,或者可能受到高温的影响,必须更换。机器区和游戏室被另外三个房间隔开,并有不同的空调系统。因此,在复赛一年后,听到加里在公开场合抱怨空调是为了深蓝的舒适度而不是他自己的舒适度而调节的,这很有趣。复赛期间,卡斯帕罗夫确实抱怨比赛室的休息区太冷。原来,他休息的地方正上方有一个通风口,后来又安装了一个气流导流面。

When we went up to the thirty-fifth floor of the Equitable Building, we found Deep Blue ensconced in a small closet-sized space at a far corner of the building. There was a dedicated air conditioner for the machine, but the space around it was quite warm. The adjacent space was a little bit cooler, and a few days later we installed a floor fan to circulate this air to help reduce the temperature. The whole machine area was too warm even for someone wearing a short-sleeve T-shirt and shorts. We did not have access to other parts of the thirty-fifth floor at the time, and it was quite difficult to work for a long period on Deep Blue, as we were forced to stay in the machine area. For a couple of days, I ran extensive tests of the system when Joe was not running his own test on the search efficiency of the parallel program. Some of the chess cards appeared to have been damaged in transport, or perhaps suffered from the effect of the high heat, and had to be replaced. The machine area and the game room were separated by three other rooms, and had different air conditioning systems. It was therefore quite funny to hear Garry complaining in public, a year after the rematch, that the air conditioner was adjusted for Deep Blue’s comfort instead of his own. During the rematch, Kasparov did complain that his resting area in the game room was too cold. It turned out that there was an air vent right above his resting area and a deflecting surface for the airflow was installed afterwards.

除了最终的测试之外,深蓝的评估功能和开本还有一些工作正在进行中。据称,我们在米开朗基罗博物馆预订了一间酒店套房,以便我们可以与大师一起创作开篇书籍。IBM 已经在酒店进行了批量预订,我们很惊讶地发现我们的套房已经被订满了。一连几天,我们除了在深蓝房间旁边的一个小房间的咖啡桌上外,没有地方可以做开业准备。最后,我们弄清楚了我们的套房发生了什么。IBM在预订酒店时,Garry想住在离中央公园较近的酒店,所以他的随行人员都订到了广场酒店。复赛期间,米开朗基罗酒店的卡斯帕罗夫营地还预留了一间套房,但之前没有。加里的新经理欧文·威廉姆斯提前一周到达米开朗基罗酒店,不小心拿走了我们预订的套房。当为卡斯帕罗夫营地预订的套房可用时,我们得到了我们的套房。

Besides the final testing, there was still some work ongoing in both the evaluation function and the opening book for Deep Blue. A hotel suite was supposedly reserved at the Michelangelo so that we could work on the opening book with the Grandmasters. IBM had made block reservations at the hotel, and we were quite surprised to find that our suite had already been taken. For several days, we had nowhere to do the opening preparation except on a coffee table in a small room next to Deep Blue’s room. Finally, we figured out what happened to our suite. When IBM was making the hotel reservations, Garry wanted to stay in a hotel closer to Central Park, so his entourage was all booked into the Plaza Hotel. A suite was also reserved for the Kasparov camp at the Michelangelo for the duration of the rematch, but not before. Garry’s new manager, Owen Williams, arrived at the Michelangelo a week early, and accidentally took the suite that we had reserved. We got our suite when the one reserved for the Kasparov camp became available.

复赛前三天,我们终于被允许搬进手术室,这里就是我们的大本营。操作室通常可能被用作编辑室,因为一侧有用于切换音轨和视频通道的面板。房间很狭窄,比赛期间通常不会超过六个人。因 UNIX 和 Belle 而出名的肯·汤普森 (Ken Thompson) 总是在场。他是两个营地的公正观察员。在比赛期间,Ken 是 IBM 之外唯一见到 Deep 的人蓝色的运行日志,除了大师和偶尔的记者。运行日志详细记录了深蓝的“思考”过程,即在各种情况下它打算做什么。三名深蓝团队成员中至少有一名始终在场,随时准备与桌子上的操作员交换位置。乔尔或米格尔总是在场,这取决于他们中的哪一个在新闻发布室。其他几个 IBM 员工时不时地进进出出。每隔一段时间,就会有各大媒体的记者被护送进去进行即兴采访。手术室里的一把椅子可能是比赛期间房子里最好的座位。游戏室和礼堂都有视频直播,因此您不会错过大楼内发生的任何重要事件。(如果能从新闻发布室获得视频就更好了,但没有可用的。)您至少有一名大师在场进行国际象棋讨论,并且至少有两名专家(肯和三名深蓝中的一两名)团队成员)进行计算机国际象棋讨论。更重要的是,你得看看什么棋最好当所有的讨论都在进行时,世界上的游戏实体正在“思考”。

Three days before the rematch, we were finally allowed to move into the operations room, which was to become our home base. The operations room was probably normally used as an editing room as there were panels for switching sound tracks and video channels on one side. The room was cramped and during the games normally no more than six people could crowd in. Ken Thompson, of UNIX and Belle fame, was always present. He was serving as the impartial observer for both the camps. During the games, Ken was the only person outside IBM who got to see Deep Blue’s running log, besides the Grandmasters and the occasional reporter. The running log detailed Deep Blue’s “thinking” process, namely, what it planned to do in various circumstances. At least one of the three Deep Blue team members was always present, ready to swap positions with the operator at the table. Either Joel or Miguel was always around, depending on which one of them was in the pressroom. A couple of other IBMers went in and out from time to time. Every once in a while, a reporter from a major media outlet would be escorted in for an impromptu interview. A chair in the operations room was probably the best seat in the house during the games. There were video feeds from the game room and from the auditorium, so you would not miss anything important happening in the building. (It would have been nice to have had a video feed from the pressroom, but none was available.) You had at least one Grandmaster on hand for chess discussion, and at least two experts (Ken and one or two of the three Deep Blue team members) for computer chess discussion. More importantly, you got to see what the best chess playing entity in the world was “thinking” while all the discussion was going on.

1997年比赛开始前,深蓝队在游戏室里。

The Deep Blue group in the game room before the start of the 1997 match.

从左到右依次为:Jerry Brody、Feng-hsiung Hsu、Joe Hoane、Joel Benjamin 和 CJ Tan。坐着:默里·坎贝尔。

Standing from left to right: Jerry Brody, Feng-hsiung Hsu, Joe Hoane, Joel Benjamin, and C J Tan. Sitting: Murray Campbell.

我们已经为演出做好了准备。

We were ready for the show.

它看到了伙伴们!

It Saw the Mates!

第一场比赛前两天,公平大厦四十九层举行了新闻发布会。原本预计大约有一百名记者到场,但最终到场的记者达到了两百多人。房间里挤满了人。ABC、CBS、NBC、CNN、福克斯、BBC 和其他一些欧洲电视网络都派出了工作人员。新闻发布会结束后,我与一位 IBM 老前辈聊天,他参与了 1964 年 IBM 360 计算机产品的发布,这可以说是 IBM 公司历史上最重要的发布。他感叹道:“这比 IBM 360 的发布规模要大得多。IBM 历史上没有任何事情可以与此相媲美。” 随着比赛的进行,越来越多的记者前来报道。最终,由于记者太多,新闻发布室搬到了第五十层的一个更大的房间,可以看到城市的壮丽景色。

Two days before game one, a press conference was held on the forty-ninth floor of the Equitable Building. It was originally estimated that about one-hundred journalists would show up but in the end over two-hundred arrived. The room was completely packed. ABC, CBS, NBC, CNN, Fox, BBC and a few other European TV networks sent their crews. After the press conference, I had a chat with an IBM old-timer who had been involved in the product launch for the IBM 360 computer in 1964, arguably the most important launch in IBM corporate history. He exclaimed, “This is far bigger than the IBM 360 launch. There is nothing in IBM history comparable to this.” As the match progressed, more and more reporters showed up to cover it. Eventually, there were simply too many reporters, and the pressroom was moved to a much bigger room on the fiftieth floor, with a spectacular view of the city.

在两场深蓝比赛中,摄影师都被允许在每场比赛开始时拍照。现在摄影师太多了,每场比赛都要分成两批。第一组被允许挤进房间,拍几张照片。然后他们被带出去,第二组被允许进入。一些摄影师被发现试图留下来进行第二次拍摄;有些摄影师被抓获。摄影师们为了拍出最好的照片,甚至发生了一些暴力事件。在第一场比赛开始时,IBM 雇用的一名保安遭到一名摄影师的殴打,随后摄影师的记者通行证被吊销。

In both Deep Blue matches, photographers were allowed to take pictures at the start of each game. Now, there were too many photographers, and they had to be divided into two batches for each game. First one group was allowed to cram into the room, and take a few pictures. Then they were ushered out and the second group allowed in. Some photographers were caught trying to stay on for the second session; there were even some cases of violence as the photographers jockeyed to get the best shots. At the start of game one, a security guard hired by IBM was punched by a photographer who subsequently had his press pass revoked.

加里和我在坐下来开始第一场比赛之前微笑着握手。复赛总奖金为110万美元,其中冠军奖金70万美元,亚军奖金40万美元。在我们握手的那一刻,他可能正在微笑着谈论他的70 万美元。我想知道当他收到 40 万美元的支票时会发生什么。我们不会分享 700,000 美元的获胜者支票——深蓝的份额将返回 IBM。但看到我们的梦想最终实现就足够了。我们都会在云九上。所有的个人牺牲都是合理的。我们中的一些人会流下喜悦的泪水吗?我可以吗?

Garry and I smiled broadly and shook hands before we sat down for the first game. The rematch had a total prize fund of $1.1 million, with $700,000 for the winner, and $400,000 for the runner-up. At the moment that we shook hands, he was probably smiling about his $700,000. I was wondering what would happen when he received his $400,000 check. We would not get to share the $700,000 winner’s check—Deep Blue’s share would go back to IBM. But seeing our dream finally coming to fruition would be more than enough. We would all be on cloud nine. All the personal sacrifices would be justified. Would some of us shed tears of joy? Would I?

尽管弗雷德里克在访问期间声称加里会获胜如果他下反计算机象棋,他会很容易,但他永远不会这么做——他就是这么做的。在整个第一局比赛中,他的棋子从未离开过他的一半棋盘。

Despite Frederic’s claim during his visit that Garry would win easily if he played anti-computer chess, but would never do it—he did just that. For the entire first game, his pieces never left his half of the board.

经过前九招之后,加里和深蓝来到了一个在宗师级别已经玩过很多次的著名位置。这个位置是封闭的,位置性很强,正是那种对计算机来说很难的位置。加里接下来下了一个不寻常的棋步,10.e3,将他的电子档案棋子向前移动一格。更正常的举动是将电子档案典当向前移动两个方格,提供典当交易。他想把这个头寸再维持一段时间。通过e3这招,他透露了整场比赛的策略——下反机棋。我们即将看到一场最高水平的最奇怪的国际象棋比赛。他为何决定采取这一策略?他为这种打法做了怎样的准备以及准备了多长时间?我有一个理论。

After the first nine moves, Garry and Deep Blue arrived at a well-known position that had been played many times at the Grandmaster level. The position was closed and highly positional, exactly the type of position that was supposedly hard for a computer. Garry next played an unusual move, 10. e3, moving his e-file pawn forward one square. A more normal move was to move the e-file pawn forward two squares, offering a pawn trade. He wanted to keep the position closed for a little bit longer. With the move e3, he revealed his strategy for the entire match—play anticomputer chess. We were about to see one of the strangest chess matches played at the highest level. Why did he decide to adopt this strategy? How and for how long did he prepare for this style of play? I have a theory.

弗雷德里克相信,如果加里下棋,他会轻松获胜。弗雷德里克和加里是亲密的朋友,所以他一定知道这个意见。我们在弗雷德里克来访期间告诉他,深蓝很有可能赢得比赛,并且可以合理地假设他将此信息转达给了他的朋友。我不知道加里在收到这个信息之前是否已经开始准备比赛了;他对与他的训练有关的事情并不特别关心。复赛开始前,他曾公开表示自己已经认真准备了。事后,他表示自己还没有做好准备。我相信他确实为复赛做了准备,至少是准备下反电脑棋。此外,我相信弗雷德里克对约克镇的访问可能引发了加里以这种方式比赛的决定。一种可能的情况是,深蓝队赛前的自信可能让他怀疑我们是否准备了一些杀手级的开场白。嗯,我们的宗师团队和深蓝一起发现了很多开局创新,但它们不一定是杀手锏3。然而,加里确实有理由保持谨慎。最简单的方法他为了避开我们的新开局,就等于完全放弃了正常的开局,来玩反电脑棋。

Frederic believed that Garry would win easily if he played anti-computer chess. Frederic and Garry were close friends, so he must have known about this opinion. We had told Frederic during his visit that Deep Blue had a good chance of winning the match and it is reasonable to assume that he relayed this information to his friend. I don’t know whether Garry had begun his match preparations before receiving this information; he was not particularly forward about things related to his training. Before the start of the rematch, he had stated publicly that he had prepared seriously for it. Afterwards, he stated that he had not prepared for it. I believe that he did prepare for the rematch, at least in the sense of preparing to play anti-computer chess. Furthermore, I believe that Frederic’s visit to Yorktown might have triggered Garry’s decision to play in this way. One possible scenario was that the prematch confidence of the Deep Blue team might have caused him to wonder whether we had prepared some killer opening lines. Well, our team of Grandmasters, together with Deep Blue, found quite a few opening innovations, but they were not necessarily killer lines3. However, Garry did have reason to be cautious. The easiest way for him to avoid our new opening lines was to play anti-computer chess by abandoning normal openings altogether.

假设加里在弗雷德里克访问约克镇后决定采用反电脑打法,那么他只有两个多月的时间来准备新的打法。由于他无法使用深蓝,商业国际象棋程序很可能是他的陪练伙伴。鉴于他在发布会上说深蓝应该以8比2击败商业项目,我猜测他对抗商业项目的成功率肯定比8比2要好。当然,他的评估是相当错误的深蓝和商业节目的相对播放强度。

Assuming that Garry decided to adopt the anti-computer style of play after Frederic’s visit to Yorktown, then he had just over two months to prepare the new style. Since he did not have access to Deep Blue, the commercial chess programs were most likely his sparring partners. Given his statement at the press conference that Deep Blue should beat the commercial programs by 8 to 2, my guess is that his success rate against the commercial program must have been better than 8 to 2. Of course, he was quite wrong in his assessment of the relative playing strength of Deep Blue and the commercial programs.

在 Garry 的 e3 之后,深蓝似乎也有同样的想法,并将其 h 档棋子移动了一格。这是一个小错误,但考虑到他的最后一步,这是可以玩的。更好的举措是培养女王。这一步棋以及深蓝的下一步,让其皇后陷入了尴尬的境地,也指出了深蓝的一个问题。当设计新的国际象棋芯片时,在国际象棋硬件上几乎没有花多少精力来设计如何定位皇后。我拥有的国际象棋书籍只对皇后的走位进行了粗略的处理。这场比赛结束后,乔写了一些简单的代码来确定皇后应该放在哪里。

After Garry’s e3, Deep Blue seemed to be in a like mind, and moved its h-file pawn one square. It was a slight error but, considering his last move, it was playable. A better move would have been to develop the queen. This move and Deep Blue’s next move, which placed its queen in an awkward position, pointed out a problem that Deep Blue had. When the new chess chip was being designed, little design effort was put in the chess hardware on how to position the queen. The chess books that I had, only gave cursory treatments on the positional play of the queen. After this game, Joe wrote some simple code on where the queen should be placed.

第 14 步后的位置。

Position after move 14.

接下来的几手,深蓝还下了一些有问题的棋,但考虑到卡斯帕罗夫自己下的不是最优的棋,深蓝的位置仍然是站得住脚的。

For the next few moves, Deep Blue played some other questionable moves, but given the fact that Kasparov was playing suboptimal moves himself, Deep Blue’s position was still tenable.

第十四步后,深蓝将位于自己王前面的 g 档棋子和 e 档棋子提升到第五位。深蓝的王几乎完全没有兵保护。f5 方格现在变成了一个弱方格——它不再能够被相邻的 e-pawn 和 g-pawn 守卫。即使在我看来,这个位置也很可怕。问题是加里能否轻松利用 f5 方格。当我仔细观察这个位置时,我突然意识到,也许深蓝对这个位置的处理还不错,尽管它肯定不是人类玩家会选择的位置。在接下来的十步左右的时间里,加里非常努力地试图将马带到f5方格,但无济于事。1996年的比赛,虽然深蓝赢了第一局,但我有一种明显的感觉,就是加里在发号施令。当我观看这场比赛时,我有一种完全不同的感觉。这感觉更像是一场谈判。他试图执行他的计划,而深蓝却不断想方设法阻止他。然后他找到了新的方法来继续他的计划,而深蓝则带着另一个反计划回来了。双方都没有得到自己想要的东西。然而,深蓝永远不会沮丧。我不知道比赛会如何结束,但我确实知道,如果加里每场下怀特的比赛都像这样,他就会陷入大麻烦。

After move fourteen, Deep Blue advanced both the g-file pawn, which was in front of its own king, and the e-file pawn to the fifth rank. Deep Blue’s king was almost completely devoid of pawn protection. The f5 square now became a weak square—it could no longer be guarded by the adjacent e-pawn and g-pawn. The position looked horrible even to my eyes. The question was whether Garry could exploit the f5 square easily. When I looked more closely at the position, it dawned on me that maybe Deep Blue’s treatment of the position was fine, even though it was certainly not one that would be chosen by a human player. For the next ten moves or so, Garry tried very hard to get a knight to the f5 square, to no avail. In the 1996 match, even though Deep Blue won the first game, I had the distinct feeling that Garry was the one calling the shots. As I watched this game unfold, I had a completely different feeling. It felt more like a negotiation. He was trying to execute his plan, and Deep Blue kept on finding moves to thwart it. He then found a new way to continue his plan, and Deep Blue came back with another counter plan. Neither side was getting what it wanted. Deep Blue, however, would never get frustrated. I had no idea how the game would end, but I did know that if every game where Garry had White was going to be like this one, he would be in deep trouble.

深蓝的棋子 22.... g4 导致其 g 档棋子与 Garry 的 h 档棋子进行交易,引起了争议。一些分析师认为这是合理的,因为深蓝获得了更好的棋子活跃度;由于加里最终没有获得明显的获胜位置,所以很难肯定地说这是一个糟糕的举动。这一举动是我们在复赛前引入的一个错误的结果。评估函数的自动调整运行向我们指出,应该增加一类国王安全项的权重。当我们增加项的权重时,深蓝似乎表现得更好,因此我们保留了新的权重。我们没有意识到,在极端情况下,新的权重达到了最大允许值,并变得饱和。换句话说,深蓝不再区分非常糟糕的位置和更糟糕的位置。因此,扔掉 g 档棋子对深蓝来说没有任何意义,只要它消除了加里的 h 棋子即可。我们是在之后才发现这个bug的第四场比赛,出现了类似的问题。在这两种情况下,都很难确定此举是否糟糕。

Deep Blue’s move 22.… g4, which resulted in its g-file pawn being traded with Garry’s h-file pawn, was controversial. Some analysts considered it justified in that Deep Blue gained better piece activity; since Garry did not get a clearly winning position as a result, it was hard to say for sure that it was a bad move. The move was the result of a bug we had introduced just before the rematch. An automatic tuning run for the evaluation function pointed out to us that the weighting for one class of king safety terms should be increased. When we increased the weights of the terms, Deep Blue appeared to play better, and we left the new weights in. What we did not realize was that, in extreme cases, the new weights reached the maximum allowed value, and became saturated. In other words, Deep Blue no longer distinguished between a very bad position, and an even worse position. Tossing away the g-file pawn therefore meant nothing to Deep Blue, as long as it eliminated Garry’s h pawn. We found the bug only after game four, when a similar problem showed up. In both cases, it was far from certain that the moves were bad.

深蓝第 22 步之后的位置。... g4。

Position after Deep Blue’s move 22.… g4.

深蓝的28.…f5这手棋将比赛带入了关键阶段。在董事会上,加里和我都很平静地对待这件事。他可能已经预料到了这一点,尽管这一步棋是在他第二十八步棋的那一瞬间下的。观众席上,解说们都惊呆了,谁都没有想到。这可能是这个位置上最好的举动,加里可以选择是否做出交换牺牲(放弃车换取主教或马)。深蓝期望他做出交换牺牲,尽管它评估牺牲后的位置对加里来说要好得多。尸检分析表明,如果他没有做出牺牲,他的处境会更糟。

Deep Blue’s move 28.… f5 put the game into a crucial phase. At the board, both Garry and I were taking it calmly. He was probably expecting it, even though the move was played at the exact instant of his twenty-eighth move. In the auditorium, the commentators were shocked—none of them expected it. It was probably the best move in the position, and Garry had the choice of whether to make an exchange sacrifice (giving up a rook for a bishop or a knight) or not. Deep Blue was expecting him to make the exchange sacrifice, although it assessed the position after the sacrifice to be much better for Garry. Post mortem analysis indicated that he would have been much worse off if he had not made the sacrifice.

深蓝第 28 步之后的位置。... f5。

Position after Deep Blue’s move 28.… f5.

事实上,深蓝并不喜欢自己的处境。通常,这意味着我会感到痛苦,但我没有——位置很乱,但我们还有机会。在礼堂和新闻发布室,人们开始谈论深蓝的意外胜利,但他们没有看到深蓝的成果。加里很快就采取了引发交换牺牲的行动。接下来的几个动作,都如深蓝所料。加里和深蓝显然都看到了所有的后果。

As it was, Deep Blue did not like its position. Usually, this meant that I would be feeling miserable, but I wasn’t—the position was messy and yet we would have our chances. In the auditorium and at the pressroom, people were starting to talk about an upset win for Deep Blue, but they were not seeing Deep Blue’s output. Garry quickly made the move that initiated the exchange sacrifice. The next few moves all went as Deep Blue expected. Both Garry and Deep Blue apparently had seen all the ramifications.

深蓝第 33 步棋之前的位置。... Qb5。

Position before Deep Blue’s move 33.… Qb5.

然而,当尘埃落定后,深蓝下了一个有问题的棋,33.…Qb5,这使得加里可以下Qfl来交换皇后。深蓝愿意交易皇后,因为它的国王不太安全。但加里还有两个相连的传球棋子(f5 棋子和 g6 棋子),如果皇后离开棋盘,它们在残局阶段可能会变得非常有价值。国际象棋硬件中存在表达这种残局特征的能力,但并未被使用。乔在第一场比赛后编写了使用这种硬件能力的代码,但直到我们在第二场比赛后休息一天检查时,新代码才处于活动状态。随后,深蓝转而打出33.…h5。

When the dust settled, however, Deep Blue played a questionable move, 33.… Qb5, which allowed Garry to play Qfl to trade off the queens. Deep Blue was willing to trade the queens because its king was less safe. But Garry also had two connected passed pawns (the f5 pawn and the g6 pawn) that could become very valuable in the endgame phase if the queens left the board. The ability to express this endgame feature existed in the chess hardware but was not in use. Joe wrote the code to use this hardware ability after game one, but the new code was not active until we had a rest day after game two to check it out. Afterwards, Deep Blue played 33.… h5 instead.

位置在 36 之前... Kf8。36.... Ng4 是最后一次抽签机会。

Position before 36.… Kf8. 36.… Ng4 was the last chance to draw.

第33步……Qb5可能很糟糕,但不一定会输,因为深蓝在第36步后面有一个听棋。而不是 36.... Kf8?,36.... Ng4!似乎在画画。深蓝的事后分析表明,如果深蓝能够再搜索一层,那么它可能会玩 36....NG4,这将需要将搜索速度提高大约六倍。

The move 33.… Qb5 was probably bad, but it was not necessarily losing, as Deep Blue had a drawing move later on move thirty-six. Instead of 36.… Kf8?, 36.… Ng4! appears to draw. Post mortem analysis with Deep Blue suggested that Deep Blue might have played 36.… Ng4 if it could have searched one more ply, which would have required roughly a six-fold increase in search speed.

没有了36....NG4,深蓝的位置很快就走下坡路。到第四十四步时,我们已经接近辞职了。然后它遇到了同样的错误,导致 Deep Blue Jr 在训练比赛中向 Christiansen 扔了一个棋子。它没有按照预期走 44.... Rf5 并延长比赛,而是走立即失败的棋 44.... Rd1,在加里回复后,我们立即辞职。那天晚上,Joe 工作到很晚,终于修复了困扰我们一个多月的 bug。

Without 36.… Ng4, Deep Blue’s position quickly went downhill. By move forty-four, we were close to resigning for it. Then it suffered the same bug that caused Deep Blue Jr to toss a pawn against Christiansen in its training match. Instead of playing 44.… Rf5 and prolonging the game as it intended, it played the instantly losing move 44.… Rd1, and after Garry’s reply, we immediately resigned. That night, Joe worked late and finally fixed the bug that had haunted us for over a month.

比赛结束后,我们知道我们还有工作要做,但全队士气高昂。我们有黑棋,但加里的反计算机国际象棋实验几乎对他产生了适得其反的效果。深蓝处理不当一旦局面变成双刃剑,我们就会占据有利位置,但是,如果这种情况继续下去,我们将作为黑方获得胜利。对于刚刚输掉六场比赛第一场比赛的一群人来说,我们在记者面前一定表现得出奇的高兴。

After the game, we knew we had work to do, but the team was in high spirits. We had Black, and yet Garry’s anti-computer chess experiment almost backfired on him. Deep Blue mishandled the position once the position became double edged but, if this scenario continued, we would have our share of wins as Black. For a group of people who had just lost the first game of a six-game match, we must have looked surprisingly cheerful to the reporters.

与此同时,加里显然对深蓝在游戏中的最后一步感到困惑。第二场比赛后的第二天是休息日,当我上网了解比赛新闻时,我看到了弗雷德里克·弗里德尔(Frederic Friedel)写的一篇关于第一场比赛后卡斯帕罗夫营地发生的事情的文章。加里对 44 棋感到困惑。... Rdl。深蓝因为一个错误而播放了它,但加里并不知道这一点。因此,整个卡斯帕罗夫阵营深入分析了为什么替代棋 44.... Rf5 不好。最后,他们得出的结论是,深蓝不玩 44....Rf5 的原因是“它可能会在 20 步或更多 [步数] 中看到交配。” 我忍不住笑了起来4

Garry, meanwhile, apparently was puzzled by Deep Blue’s last move in the game. The day after game two was a rest day, and when I surfed the web to catch up on the match news, I came across an article written by Frederic Friedel about what happened in the Kasparov camp after game one. Garry was perplexed by the move 44.… Rdl. Deep Blue played it as a result of a bug but Garry did not know that. So the whole Kasparov camp went into a very deep analysis on why the alternative move 44.… Rf5 was no good. In the end, they concluded that the reason why Deep Blue did not play 44.… Rf5 was “It probably saw mates in twenty or more [moves].” I could not help but burst out laughing4.

坏招致胜

A Bad Move That Wins

第二场比赛前一天早上,团队讨论是否应该改变深蓝的评估功能。乔已经修复了第一场比赛中最后一步的错误,但我们还需要考虑其他两个更改。对已通过的棋子评估的拟议更改太过剧烈,我们决定将其保留到第二场比赛中。乔对女王评价的改动相对较小,我们决定将其保留在程序中。玩第二场游戏的程序本质上与玩第一场游戏的程序相同。

The morning before game two, the team debated whether we should make changes to Deep Blue’s evaluation function. Joe had already fixed the bug responsible for the last move in game one, but there were two other changes that we needed to look at. The proposed change to the passed pawn evaluation was too drastic, and we decided to leave it out for game two. Joe’s change to the queen evaluation was relatively minor, and we decided to keep it in the program. The program that played game two was essentially the same program that played game one.

穆雷是当天的首发操作员,所以我在比赛的早期阶段就在操作室里度过了一段时间。深蓝以 1.e4 开局。加里没有用正常的西西里语 1.... c5 进行回复,而是打出了 1.... e5,就像他在 1996 年比赛的第五场比赛中所做的那样。如果 Garry 玩 1.…c5,我们就准备好用西西里岛的主线来与他战斗,而不是 1996 年玩的 c3 变体。

Murray was the starting operator for the day, so I spent time in the operations room during the early part of the game. Deep Blue opened with 1. e4. Garry, instead of replying with his normal Sicilian 1.… c5, played 1.… e5, just as he did in game five of the 1996 match. Had Garry played 1.… c5, we would have been ready to battle him with the main lines of the Sicilian instead of the c3 variation that we played in 1996.

该游戏很快成为鲁伊·洛佩兹开局(也称为西班牙酷刑),以 15 世纪西班牙牧师的名字命名。尼克和乔尔,而不是我们的西班牙队伍米格尔,为鲁伊·洛佩兹准备了基本台词。随着开局的迅速展开,乔尔的心情异常乐观。他曾经玩过鲁伊·洛佩兹在与深蓝Jr的比赛中打了很多场比赛,对深蓝处理这个位置的能力很有信心。

The game quickly became a Ruy Lopez opening (also known as the Spanish Torture), named after a 15th-century Spanish priest. Nick and Joel, rather than Miguel our Spanish contingent, had prepared the basic lines for the Ruy Lopez. As the opening quickly unfolded, Joel was in an unusually upbeat mood. He had played many games in the Ruy Lopez against Deep Blue Jr and was quite confident of Deep Blue’s ability to handle the position.

与此同时,观众席上,解说们纷纷议论深蓝的开场准备有多糟糕。深蓝让加里达到了封闭的位置。甚至在比赛开始阶段结束之前,评论员就已经取消了“深蓝”的称号。虽然没有明说,但很明显,他们已经预料到深蓝会输掉比赛。

Meanwhile, in the auditorium, the commentators were saying how terrible Deep Blue’s opening preparation was. Deep Blue had allowed Garry to achieve a closed position. The commentators had written off Deep Blue even before the opening phase of the game was completed. It was not said, but it was clear that they were expecting Deep Blue to lose the game.

深蓝棋子16.d5后的位置。

Position after Deep Blue’s move 16. d5.

当深蓝打出16.d5时,加里可能很高兴,封锁了中锋,表面上给了他他所渴望的封闭位置。乔尔也很高兴。加里的开局选择给深蓝带来了巨大的空间优势,这可能会在剩下的比赛中困扰加里。乔尔曾是许多具有相同主题的惩罚性游戏的受害者。在作战室里,我们缓慢而坚定地看到深蓝对其位置优势的评估正在悄然上升。礼堂里的解说员继续说深蓝有麻烦了。大师Yasser Seirawan:“这不是电脑能打得很好的局面。” 国际大师迈克·瓦尔沃(Mike Valvo):“从计算机的角度来看,这种位置并没有什么真正清晰的地方。”

Garry was probably quite happy when Deep Blue played 16. d5, blocking up the center, ostensibly giving him the closed position that he so desired. Joel was also quite happy. Garry’s opening selection had given Deep Blue a significant advantage in space that could haunt Garry for the rest of the game. Joel had been on the receiving end of many punishing games with the same theme. In the operations room, slowly but surely we saw Deep Blue’s assessment of its positional edge creeping up. The commentators in the auditorium continued to say that Deep Blue was in trouble. Grandmaster Yasser Seirawan: “This isn’t the kind of position that the computer can play very well.” International Master Mike Valvo: “From the computer’s point of view, there’s nothing really clear in this kind of position.”

乔尔可能感觉到有机会刺杀他认识多年的亚西尔,请求离开手术室,以便他可以到礼堂台为深蓝的开场准备辩护。几乎在他离开后,评论员们有了更多的时间来消化董事会的立场,并看到加里轻轻摇头,他们改变了主意。亚西尔开始说他喜欢怀特。

Joel, possibly sensing a chance to needle Yasser, whom he had known for years, asked to be excused from the operations room so that he could get to the auditorium stage to defend Deep Blue’s opening preparation. Almost immediately after he left, the commentators, having had more time to digest the board position and seeing Garry shaking his head a little bit, had a change of heart. Yasser was starting to say that he liked White.

深蓝移动后的位置 24. Rec1。

Position after Deep Blue’s move 24. Rec1.

深蓝的“类人”着法24.Rec1迫使对手做出艰难的决定。加里的24.... c4关闭c档在赛后被批评为消除了他自己最后的反击机会。深蓝已经准备好自己打c4并将整个位置打开,而Garry的举动至少阻止了这一点。加里可能不想让事情变得复杂,考虑到深蓝拥有更活跃的棋子,这是一个可以理解的决定。

Deep Blue’s “human-like” move 24. Rec1 forced its opponent to make a hard decision. Garry’s 24.… c4 closing up the c file was criticized after the game as removing his own last chance of counter play. Deep Blue was ready to play c4 itself and bust the whole position wide open and Garry’s move at least stopped that. Garry probably did not want to allow the complications and, given the fact that Deep Blue had the more active pieces, it was an understandable decision.

乔尔此时到达礼堂舞台,对亚西尔进行针刺有点太晚了。乔尔的第一句话是:“好吧,我在楼上听你的评论,你在谈论这个职位长期以来一直处于封闭状态。我想知道你们什么时候才能意识到这个位置对怀特来说更好,而你们最终做到了。” 观众笑了。

Joel reached the auditorium stage at this point, a little too late to needle Yasser. Joel’s first remarks were, “Well, I was following your commentary from upstairs, and you were talking about how closed the position was for so long. I was wondering when you guys were going to get around to the fact that the position was better for White, and you eventually did.” The audience laughed.

在操作室里,当深蓝将车放在 a 档上时,乔和我交换了一个眼神,但拒绝立即打开 a 档。我们俩都在思考几个月前乔敲我的门后发生的事情。关于潜在开放文件上的车的新评估功能在这里发挥作用。我微笑着对乔说:“对于一个下午的工作来说还不错。” 潜在的空档上的两辆白车将加里的重型棋子绑在他的后排,迫使他完全被动。

In the operations room, Joe and I exchanged a look when Deep Blue doubled up its rooks on the a-file, and yet refused to open up the a-file immediately. Both of us were thinking about what happened many months ago after Joe knocked on my door. The new evaluation feature about rooks on potential open files was in play here. I smiled and said to Joe, “Not bad for an afternoon’s work.” The two White rooks on the potential open a-file tied down Garry’s heavy pieces to his back rank, and forced him to be completely passive.

深蓝移动26.f4后的位置。

Position after Deep Blue’s move 26. f4.

深蓝的26.f4,在王方开辟了第二条战线,可能部分是新评估功能的结果。阿纳托利·卡尔波夫正在一家互联网国际象棋俱乐部现场观看比赛。他在比赛初期发表了很多评论,并且对深蓝的表现相当不屑一顾。深蓝的26.f4也没有逃过他的批评;尽管据我所知,其他大师都没有同意他的观点——事实上,他们普遍赞扬了这一举动。随着比赛的进行,阿纳托利的言论越来越少,最后陷入了沉默。

Deep Blue’s 26. f4, opening up a second front on the king side, was probably partially a result of the new evaluation feature. Anatoly Karpov was watching the game live in one of the Internet chess clubs. He was making quite a few comments early in the game, and was fairly dismissive of Deep Blue’s play. Deep Blue’s 26. f4 did not escape his criticism; although, to the best of my knowledge, none of the other Grandmasters shared his view—in fact, they universally praised the move. As the game progressed, Anatoly made fewer and fewer comments, and towards the end fell silent.

在游戏室里,穆雷示意他需要替换。加里还在棋盘上思考,所以我沿着漫长曲折的路线穿过正常的入口进入游戏室。我和其他观众一起坐下来,等待加里的举动,然后才与穆雷进行交换。当我在手术室时,加里在视频监视器上看起来很不高兴,但其他方面很正常。当我在游戏室里看到他时,我吃了一惊。他看起来就像是刚从一场灾难中醒来的人恶梦。他的脸有一部分明显红了,好像这一侧睡了一段时间。

In the game room, Murray signaled that he needed a replacement. Garry was still at the board thinking, so I took the long tortuous route through the normal entrance into the game room. I sat down with the other spectators and waited for Garry’s move before making the switch with Murray. When I was in the operations room, Garry looked unhappy but otherwise normal on the video monitor. I was taken aback when I looked at him in the game room. He looked like someone who had just woken up from a bad nightmare. Part of his face was visibly red as if he had slept on that side for a while.

在我接任操作员后不久,深蓝就下了两步棋,这让加里在接下来的比赛中感到非常悲伤。它们是 36.axb5 和 37.Be4。就在 axb5 之前,它下出了 Bxd6 棋,随后是 Garry 自己的 Bxd6。在计算Bxd6这一步时,深蓝的评价下降了。结果,它进入了恐慌模式,最终花了不到 15 分钟的时间走到了第 35 步。正如您稍后将看到的,Bxd6(这是一个明显的棋步)消耗的计算时间比接下来两步棋的总和还多,尽管加里后来声称深蓝在 36.axb5 上花费的时间比比赛中任何其他棋步都多。(他可能指的是 37.Be4,但深蓝只花了正常的时间。)

Soon after I took over as the operator, Deep Blue played the two moves that would cause Garry a lot of grief for the rest of the match. They were the moves 36. axb5 and 37. Be4. Right before axb5, it played the move Bxd6, which was followed by Garry’s own Bxd6. While computing for the move Bxd6, Deep Blue’s evaluation dropped. As a result, it went into the panic mode and ended up spending just slightly less than 15 minutes on its move thirty-five. As you will see later, Bxd6, which was an obvious move, used up more computation time than the next two moves combined, despite Garry’s later claim that Deep Blue spent more time on 36. axb5 than any other move played in the match. (He might have meant 37. Be4, but Deep Blue spent only a normal amount of time on it.)

在写上一段的时候,我自比赛以来第一次看到 Bxd6 着法的 log 5 ,它显示了一些非常有趣的东西。

While writing the previous paragraph, I saw the log5 for the move Bxd6 for the first time since the match itself, and it showed something quite interesting.

深蓝第 36 步棋之前的位置。axb5。

Position before Deep Blue’s move 36. axb5.

首先,当深蓝完成对 Bxd6 的恐慌模式搜索时,它计划玩 36.Qb6 作为 Bxd6 的后续游戏,但只是勉强进行,因为它已经看到了 Garry 可能会做一些令人讨厌的事情的暗示。

First, when Deep Blue finished the panic mode search for Bxd6, it was planning to play 36. Qb6 as the follow-up to Bxd6, but only barely, because it already saw hints that Garry might be able to do something nasty.

在深蓝进入 Bxd6 的恐慌模式之前,它在完成 9 层搜索时提出了一个不寻常的主要变体。它给出的主线有 15 层深。然而,这个数字只是主工作站上的搜索深度,并不包括从工作站(2+层)和国际象棋筹码(4+层)上的额外搜索深度。主线实际上至少有 21 层深,甚至可能超过 30 层深,因为工作站和芯片上还有其他搜索扩展。深蓝预计加里会放弃两个棋子并推动他的电子文件棋子以进行潜在的攻击。然而,这个分数仅略高于深蓝的一个典当,这表明它认为加里的位置补偿将略低于一个典当。

Before Deep Blue went into the panic mode for Bxd6, it had come up with an unusual principal variation when it finished its 9-ply search. The main line it gave was 15-ply deep. This number, however, was only the search depth on the master workstation, and did not include the additional search depths on the slave workstations (2+ plies), and on the chess chips (4+ plies). The main line was really at least 21-ply deep and maybe well over 30-ply deep, since there were additional search extensions on the workstations and on the chips. Deep Blue was expecting Garry to give up two pawns and push his e-file pawn for a potential attack. The score was, however, only slightly more than a pawn up for Deep Blue, signifying that it thought Garry would have slightly less than a pawn worth of positional compensation.

在 10 层搜索结束时,它预计 Garry 会牺牲三个棋子,但它的得分进一步下降到只有一个棋子。第十一层时,Bxd6的分数再次下降,导致深蓝进入恐慌模式。主要的变化再次是加里牺牲了三个棋子,但现在显示他在此​​过程中重新夺回了一个棋子。它给出的线现在有 18 层深,或者实际上至少有 24 层深,甚至沿着最深的线可能超过 40 层深。我记得在比赛过程中我对自己说:“也许 Be4 阻止电子文件棋子在这附近是一个不错的举动。”

At the end of its 10-ply search, it was expecting Garry to sacrifice three pawns, but its score had dropped further to only a pawn up. At the eleventh ply, the score for Bxd6 dropped again, which then triggered Deep Blue to go into the panic mode. The principal variation once again had the three-pawn sacrifice by Garry, but it now showed him recapturing a pawn in the process. The line it gave was now 18-ply deep, or was really at least 24-ply deep and maybe even over 40-ply deep along the deepest line. I remembered that I was saying to myself at this point during the game, “Perhaps Be4 blocking the e-file pawn would have been a good move somewhere around here.”

与此同时,在我不知情的情况下,亚西尔在礼堂和互联网上向观众提供了两兵牺牲台词。深蓝在其 9 层搜索过程中发现了相同的牺牲线,但此后将其作为黑方的劣线而拒绝。

Meanwhile, unbeknownst to me, Yasser gave the two-pawn sacrifice line to the audience at the auditorium and over the Internet. Deep Blue had discovered the same sacrifice line during its 9-ply search but had since rejected it as an inferior line for Black.

加里下棋 35 后,世界上所有其他计算机和礼堂里的观众都期待深蓝下 36.Qb6,并在此过程中赢得棋子。但现在深蓝更加不喜欢黑棋对Qb6所能得到的补偿。这次它以不同的方式进入恐慌模式。当 11 层搜索开始时其正常搜索时间结束时,预期的主变分在搜索深处的某个地方崩溃了。它进入了暂时的恐慌模式,迫使它解决移动分数Qb6. 当这次搜索完成时,分数已经下降到48分(略低于半个棋子)6。由于这个分数比预期分数74低了15分以上,深蓝继续处于恐慌状态,检查替代动作。接下来完成的棋子是axb5,分数为63。这个分数已经足够好了,所以深蓝立即下出36.axb5,棋子花费了大约6.5分钟。下一步仍打算下Qb6。Garry 用 axb5 回复。

After Garry made his move thirty-five, all the other computers in the world and the audience in the auditorium were expecting Deep Blue to play 36. Qb6, winning a pawn in the process. But now Deep Blue even more disliked the compensation that Black could get against Qb6. This time it went into the panic mode in a different way. When its normal search time was up at the start of the 11-ply search, the expected principal variation was falling apart somewhere deep down in the search. It entered a temporary panic mode that forced it to resolve the score for move Qb6. When this search was completed, the score had dropped to 48 points (slight less than half a pawn)6. Since this score was more than 15 points lower than the expected score 74, Deep Blue continued to stay in panic mode to examine the alternative moves. The next move completed was axb5, which had a score of 63. This score was good enough, so Deep Blue immediately played 36. axb5, spending about 6.5 minutes on the move. It was still planning to play Qb6 on the next move. Garry replied with axb5.

在第 37 步棋中,第 11 层 Qb6 的分数下降到 32 分,搜索进行到 84 秒。搜索进行到126秒后,Be4得到37分,取代Qb6成为最佳棋步。11 层搜索在 182 秒后完成。搜索在 199 秒后按正常时间终止。深蓝在第 37 步时并没有进入恐慌模式,它只是度过了正常的时间分配。正如前面提到的,正如后来的事件所表明的那样,最后一个事实将变得有趣。

On the thirty-seventh move, the score for Qb6 at the eleventh ply dropped to 32 points, 84 seconds into the search. After 126 seconds into the search, Be4 was scored as 37 points and replaced Qb6 as the best move. The 11-ply search was completed after 182 seconds. The search was terminated at its normal time after 199 seconds. Deep Blue did not go into panic mode on its 37th move, it just spent its normal allocation of time. As mentioned earlier, this last fact would become interesting, as later events showed.

很难确切地说是什么导致一台每秒搜索两亿个位置的国际象棋机器下出特定的棋步。对于所讨论的举措,我认为有两件事发挥了作用。深蓝沿着临界线进行了非常深入的搜索,其王者安全评价的权重比其他程序要大得多。

It is hard to say with any certainty what causes a chess machine that searches two-hundred million positions per second to play a particular move. For the moves in question, I believe that two things came into play. Deep Blue was searching very deep along the critical lines, and its king safety evaluation had much greater weight than in other programs.

对于36.axb5,做出这个决定是因为非常深入的计算。事实证明,36.Qb6 客观上可能是更好的棋步,因为白棋保留了回到类似于游戏继续的位置的选择,同时给了黑棋立即失败的机会。根据Garry的赛后评论,Qb6有可能几乎立即结束比赛,因为他没有看到最强的路线。但那样我们就不会看到第二场比赛非常有趣的结局了。

For 36. axb5, the decision was made because of very deep calculations. It turns out that 36. Qb6 might be objectively the better move, as White reserves the option of getting back into positions similar to the game continuation, while giving Black a chance to lose immediately. Based on the post game comments given by Garry, it was possible that Qb6 would have ended the game almost instantly since he did not see the strongest line. But then we would not have the very interesting finish of game two.

有证据表明,37 Be4 棋的走法决策中涉及了王的安全评估,这更多的是位置决策,而不是战术决策。第四场比赛结束后,我们发现并修复了第一场比赛早些时候出现的国王安全漏洞。在 Joe 修复了第一个错误后,深蓝停止玩 Be4,这有效地缓和了国王安全项。当它再次出现时,深蓝再次演奏了Be4。由于每个人似乎都认为 Be4 是一个伟大的举动,乔决定使用第二个错误修复,这使得国王的安全评估大致回到与第二场比赛相同的水平,第五场和第六场。

There was some evidence that the king safety evaluation was involved in the move decision for move 37 Be4, which was more a positional than a tactical decision. After game four, we found and fixed the king safety bug that occurred earlier in game one. Deep Blue stopped playing Be4 after Joe made the first bug fix, which effectively toned down the king safety term. When it was turned up again, Deep Blue again played Be4. Since everyone seemed to think that Be4 was a great move, Joe decided to use the second bug fix, which had the king safety evaluation roughly back at the same level as in game two, for games five and six.

我对 Be4 比赛时的情况不太满意。深蓝评估该位置仅具有轻微优势。幸好加里似乎看不懂我的表情。当加里接下来的几个动作中,深蓝的评价开始慢慢上升时,我的心情有所好转。我原以为这场比赛会持续很长一段时间。然后,在深蓝第四十五步之后,加里突然伸出手说道:“我辞职。” 我很惊讶,因为现在辞职似乎有点早,但他显然已经受够了西班牙式的折磨一天了。深蓝的排名领先了约160分。这是一个胜利的优势,但比赛还没有真正结束。加里可能以乐观的心情开始比赛,尤其是当深蓝封闭中锋时。从那时起,他就被迫采取防守动作。压力越来越大,他终于无法承受了。

I was not really too happy about the situation when Be4 was played. Deep Blue assessed the position as only slightly advantageous. It was a good thing that Garry did not seem able to read my face. My mood improved somewhat when, over Garry’s next few moves, Deep Blue’s evaluation started to creep upward. I was expecting the game to last quite a while. Then, suddenly, after Deep Blue’s forty-fifth move, Garry offered his hand and said, “I resign.” I was surprised as it seemed a little early to resign, but he apparently had had enough Spanish torture for one day. Deep Blue scored the position as about 160 points up. It was a winning edge, but the game was not really over. Garry probably started the game in an optimistic mood, especially when Deep Blue closed up the center. From then on, he had been forced to play defensive moves. The pressure just kept on building, and finally he could not take it any longer.

加里辞职后没有再说什么。这次没有进行事后讨论。他带着随从迅速离开了大楼。我们在热烈的掌声中走上礼堂舞台。我们兴高采烈。我们去年的努力终于得到了世界的认可。默里、乔和我过去时常争论,但在过去一年左右的时间里,我们抛开了分歧,为共同的目标而努力。我们和大师们一起成为了一个真正的团队,所有的团队合作终于得到了回报。就个人而言,这是一个非常令人满意的时刻。我在新的国际象棋芯片上花费的六个月的工作是完全值得的。在那六个月里,我全身心投入到这个项目中。整整六个月没有周末,深夜开车回家,每晚醒来几次检查模拟,这一切突然变得值得了。我们很高兴。我们感觉就像是自豪的父母,刚刚看到他们的新生婴儿迈出了重要的第一步。

Garry did not say anything else after he resigned; there was no post-mortem discussion this time. He left the building quickly with his entourage. We went on to the auditorium stage to a round of applause. We were exuberant. Our labor for the last year had finally been recognized by the world. Murray, Joe and I had argued from time to time in the past, but for the last year or so we had put aside our differences and worked on the common goal. We had come together with the Grandmasters to be a real team, and all the teamwork had finally paid off. Personally, it was a very satisfying moment. The six months of work that I had spent on the new chess chip was completely justified. For those six months, I had poured my body and soul into the project. Having no weekends for six whole months, driving home late at night and waking up several times a night to check the simulation had all suddenly become worthwhile. We were happy. We felt like proud parents who had just seen their new baby taking the first significant step.

亚西尔在第二场比赛开始时并没有给予深蓝太多的信任,但到最后他改变了主意:“我会为拥有这个而感到自豪。” 莫里斯·阿什利:“这是一场精彩的比赛。” 乔尔·本杰明 (Joel Benjamin) 说道:“这是一场任何人类大师都会为自己为白棋而感到自豪的游戏。这不是电脑类型的游戏。这是真正的国际象棋。” 这是有史以来最好的电脑游戏。

Yasser was not giving Deep Blue much credit at the beginning of Game two but by the end he changed his mind: “I would be proud to have this one.” Maurice Ashley: “It’s a gorgeous game.” Joel Benjamin had the following to say: “This is a game that any human grandmaster would be proud to have played for White. This was not a computer-type game. This was real chess.” It was the best game ever played by a computer.

然而,有一个美中不足的地方。

Yet, there was a fly in the ointment.

第二天我起了个大早,正好是休息日。还沉浸在上一场比赛的兴奋之中,我走进指挥室查看网络新闻。我访问的网站之一是 Mark Crowther 的“The Week In Chess”。几年前我和马克有过一些交流。他认为计算机会毁掉国际象棋,而国际象棋世界冠军输给计算机对国际象棋来说将是非常糟糕的。我认为人机国际象棋比赛可能是国际象棋的一大福音,国际象棋世界冠军最终输给电脑实际上对国际象棋界来说是一个很好的机会。巨大的宣传会吸引人们参与这项运动,如果国际象棋界能够积极响应,利用宣传向新的受众宣传国际象棋的乐趣和享受,那么这项运动的受欢迎程度可能会大幅提高。马克并不完全相信。也许他同意我的逻辑,但不确定国际象棋界会如何反应。

I woke up early the next day, which happened to be a rest day. Still excited from the last game, I went into the operations room to catch up on the Internet news. One of the sites that I visited was Mark Crowther’s “The Week In Chess”. I had had some communication with Mark a few years back. He felt that computers would destroy chess, and a loss by the World Chess Champion to a computer would be very bad for the game. I suggested that man-machine chess matches could be a great boon to chess, and the eventual loss by the World Chess Champion to a computer would actually be a great opportunity for the chess world. The huge publicity would draw people to the game, and if the chess world could respond positively and take advantage of the publicity to promote the fun and enjoyment of chess to the new audience, there could be a massive increase in the game’s popularity. Mark was not completely convinced. Perhaps he agreed with my logic but was not sure about how the chess world would react.

马克网站上关于第二场比赛的第一个声明让我感到惊讶。他明确表示第二场比赛应该是平局。我的第一反应是,“马克否认了。” 我粗略地浏览了马克对第二场比赛最终位置的分析,然后继续阅读其他新闻,然后回到酒店。早餐时,我向穆雷提到马克声称第二场比赛应该是平局。在与默里交谈时,我想到了一个想法:“如果我是否认的人怎么办?” 早餐后,我决定仔细检查网上给出的分析。

The first statement on Mark’s web site about game two surprised me. He made a blanket statement that game two should be a draw. My first reaction was, “Mark is in denial.” I took a cursory glance at Mark’s analysis of the final position in game two, and went on to other news items before I went back to the hotel. Over breakfast, I mentioned to Murray that Mark was claiming that game two should be a draw. While talking with Murray, a thought occurred to me, “What if I am the one in denial?” I decided to double-check the analysis given on the web after breakfast.

我回到操作室并与 Deep Blue Jr 一起尝试了网络上给出的变化,因为 Deep Blue 被用来测试我们的开局准备工作。不到二十分钟,我就震惊了。是的,看起来这个位置可能是加里表现最好的平局。白棋有一种方法可以避免彻底重复平局,但最终的结局尽管对白棋来说更好,但可能不会获胜。深蓝在比赛中的最后两步结果被证明是错误的。第 44 步。Kf1,响应 Garry 的 Qb6+,将王从 g1 移动到 f1,尽管仍然获胜,但让生活变得更加艰难。更好的棋步是 44.Kh1,这手棋很容易获胜,因为之后黑方没有机会重复平局。深蓝下棋44.Kf1,因为它认为这可以迫使皇后离开棋盘,而白王更靠近中心,结局会更好。真正的罪魁祸首是游戏的最后一步棋,45.Ra6,它让黑方要么获得稍差的结局,要么获得非常深的重复平局。后来我发现,这条画线最初是由人类国际象棋棋手和计算机国际象棋程序共同努力发现的。黑方最好的棋是45....Qe3,从人类的角度来看这是相当容易看到的,但是黑方的一些后续棋对人类来说并不那么容易。加里显然没有看到后续动作,因此辞职。

I went back to the operations room and tried out the variations given on the web with Deep Blue Jr, as Deep Blue was being used to test out our opening preparations. Within twenty minutes, I was in shock. Yes, it looked like the position might be a draw with best play by Garry. There was a way for White to avoid an outright repetition draw, but the resulting ending, even though better for White, might not be winning. Deep Blue’s last two moves in the game turned out to be mistakes. Move 44. Kf1, moving king from g1 to f1 in response to Garry’s Qb6+, even though still winning, made life tougher. A better move would have been 44. Kh1, which wins easily as there was no chance of a repetition draw for Black afterwards. Deep Blue played 44. Kf1 because it thought that it could force the queens off the board, and the ending would have been better with the White king closer to the center. The real culprit was the last move of the game, 45. Ra6, which allowed Black to either get a slightly inferior ending or a very deep repetition draw. I later discovered that the drawing line was first found by painstaking group efforts involving both human chess players and computer chess programs. Black’s best move is 45.… Qe3, which is fairly easy to see from a human perspective, but some of the follow-up moves for Black are not so easy for human beings. Garry obviously did not see the follow-up moves and hence the resignation.

第 2 场比赛的最终位置。

The final position in game 2.

45....Qe3 之后,白方可以下 46.Qd7+ 或 46.Qxd6。为了反驳任何一个举动,你需要进行相当深入和棘手的分析。《深蓝小号》在几分钟之内就完成了主线的关键动作,但要解决所有的后果却花了近一个小时。最深的线大约有30层深。理论上,深蓝应该能够实时覆盖30层强迫线。但有一个巨大的并发症。深蓝,像大多数其他国际象棋程序,使用了所谓的换位表。当新位置是已搜索的旧位置的重复时,换位表允许国际象棋程序避免搜索“相同”位置两次。问题在于,新职位与旧职位并不完全相同。他们有不同的搬家历史。通常,这一事实并不重要,但当涉及重复时,程序可能会由于使用换位表而错过重复或找到错误的重复。有时,当程序进行更深入的搜索时,问题最终会消失,而有时,问题永远不会消失。不管怎样,要么深蓝搜索得不够深入,要么是因为换位表而无法看到潜在的重复抽签。

After 45.… Qe3, White could play either 46. Qd7+ or 46. Qxd6. To refute either of the moves, you need to do fairly deep and tricky analyses. Deep Blue Jr had the critical moves in its main line within minutes, but to sort out all the ramifications took close to an hour. The deepest line was about 30 plies deep. Deep Blue, in theory, should be able to cover 30 plies of forcing lines in real time. But there was one huge complication. Deep Blue, like most other chess programs, used the so-called transposition table. The transposition table allows a chess program to avoid searching the “same” position twice when the new position is a repeat of an old position already searched. The problem is that the new position is not really the same as the old one. They have different move histories. Normally, this fact does not matter, but when repetitions are involved, a program could either miss a repetition or find a false repetition because of the use of the transposition table. Sometimes, when the program searches deeper, the problem eventually goes away, and sometimes, the problem never goes away. Anyway, either Deep Blue was not searching deep enough, or it was sidetracked by the transposition table from seeing the potential repetition draw.

从纯粹的国际象棋角度来看,深蓝第二局的最后一步Ra6值得用“??”来标注。标记(这意味着非常可疑的举动),因为它可能会让胜利变成平局。不过,从心理学的角度来看,此举值得用“??!!”来注解。标记(!! 意味着一个出色的举动。??!! 的意思是“一个非常可疑但出色的举动”)。它确实赢得了比赛,而根据随后比赛的情况,甚至有人会说,此举赢得了比赛。

From the pure chess point of view, Deep Blue’s last move in game 2, Ra6, deserves to be annotated with a “??” mark (this means a highly dubious move), as it could have allowed a win to be turned into a draw. However, from the psychological point of view, the move deserves to be annotated with a “??!!” mark (!! means an excellent move. ??!! would mean something like “a highly dubious but excellent move”). It did win the game, and in light of what happened subsequently in the match, some might even say that the move won the match.

茶杯里的风暴

Tempest in a Teacup

当我意识到第二场比赛可能会打平时,我给穆雷打电话。我们都感到失望。另一方面,胜利就是胜利,看看加里将如何回应这个消息将是非常有趣的。我知道我不想成为告诉他的人。至于深蓝团队,我们过得很轻松。我们的硅基播放器对此毫不在意。

I phoned Murray the moment I realized that game two could have been a draw. We both felt a letdown. On the other hand, a win is a win, and it would be quite interesting to see how Garry would respond to the news. I knew that I would not want to be the one to tell him. As for the Deep Blue team, we had it easy. Our silicon-based player could not care less.

弗雷德里克·弗里德尔 (Frederic Friedel) 写了一篇文章,介绍卡斯帕罗夫阵营如何发现潜在的划线以及加里如何回应这一消息。弗雷德里克(Frederic)很喜欢戏剧性,所以我总是对他说的话持保留态度,尤其是当它与他出售的商业国际象棋程序弗里茨(Fritz)有关时。不管怎样,他的叙述确实提出了一些其他报道所遗漏的事情。

Frederic Friedel wrote an article on how the Kasparov camp found out about the potential drawing lines and how Garry responded to the news. Frederic had a flare for the dramatic so I always took what he said with a pinch of salt, especially when it was related to Fritz, a commercial chess program that he sold. Anyway, his account did suggest a few things that were missed by other reports.

根据这篇文章,弗雷德里克在第二场比赛后工作到很晚,让弗里茨沿着潜在的抽签线走下去并分析最终的位置。他在弗里茨的帮助下找到了绘画动作,但不知道上面已经发生了什么互联网。不过,此时他还不太确定这个位置是否真的是平局。这并不令人意外。导致重复抽牌的线条太深了,毫不奇怪,弗里茨看不到从第二场比赛的最终位置7开始的重复。

According to the article, Frederic worked late after game two using Fritz to go down the potential drawing line and analyze the resulting positions. He found the drawing moves with the help of Fritz, but without knowing what had already transpired over the Internet. However, at this point he was not too sure whether the position was indeed a draw. This was not a big surprise. The lines that lead to repetition draws were simply too deep, and Fritz, not surprisingly, could not see the repetitions starting from the final position of game two7.

第二天早上,他邀请加里的第二场比赛尤里·多霍基安(Yuri Dohokian)到他的房间看分析。Frederic:“尤里在 46 时急促地吸了一口气……Re8 说,‘我们没有考虑这一步棋。’ 他开始检查我的分析,过了一段时间后得出了令人震惊的结论:这一举动确实有吸引力。” 从上述陈述中并不清楚尤里的“我们”是否包括加里本人。但只要说加里在比赛中至少错过了一个关键的绘图动作,并且可能在他自己的事后分析中,当时他有足够的时间,那就足够了。

The next morning, he asked Yuri Dohokian, Garry’s match second, to come to his room to look at the analysis. Frederic: “Yuri drew a sharp breath at 46.… Re8 and said, ‘We didn’t look at this move.’ He started checking my analysis and after some time came to the shocking conclusion: this move really draws.” It is not clear from the above statement whether Yuri’s “we” included Garry himself. But suffice it to say that Garry missed at least one critical drawing move during the game and possibly during his own postmortem analysis when he had all the time that he needed.

关于加里的反应,弗雷德里克写道:“加里捂着头,僵在第五大道中央。没有咒骂,没有咒骂,只有令人震惊的沉默。” 他们正在去一家意大利餐厅的路上,弗雷德里克描述了餐厅里发生的事情,“……这顿饭一定很糟糕。事实并非如此。他很棒。大约十到十五分钟后,他突然兴奋起来,开始给我们讲一些俄罗斯电影。” 因此,根据弗雷德里克的说法,加里对这个消息反应良好。

Of Garry’s reaction, Frederic wrote, “Garry clutched his head and froze in the middle of the Fifth Avenue. There were no expletives, no cursing, just stunned silence.” They were on their way to an Italian restaurant and Frederic described what happened in the restaurant, “… it must have been a miserable meal. It wasn’t. And he’s amazing. After about ten or fifteen minutes, he suddenly was fired up and started telling us about some Russian movie.” So, according to Frederic, Garry took the news well.

虽然可能平局的消息在卡斯帕罗夫阵营引发了危机,但在深蓝队员中,第三局比赛前的主要话题是加里是否会继续下反电脑棋。我们的第一感觉是他会开始下正常的棋,因为他的反电脑棋战术做得不太好。他确实赢得了第一场比赛,但深蓝本可以在比赛临近结束时确保平局。从我们的角度来看,Garry 有点幸运,深蓝没有进一步混淆,甚至可能赢得了比赛。第二场比赛对他来说简直就是一场灾难。他也可以确保平局,但就像第一场比赛中的深蓝一样,他也没有看到平局。无论如何,加里第二场比赛失败的根源显然是他想要玩反计算机国际象棋,这导致他达到了可怕的位置,这些位置本应对其他计算机有利,但结果对深蓝却很糟糕。但这是基于分析纯粹的国际象棋优点。我们对加里精神状态的业余心理分析得出了截然不同的结论。他有非常坚强的意志;大多数人甚至会说他有时很固执。他带着既定的计划来参加比赛,显然就是要下反电脑棋。除非被证明彻底失败,否则他很可能不会改变自己的计划。一个很好的问题是,现在我们回顾过去,如果他在整个复赛中都正常下棋,会发生什么?我不知道答案,但我相信这也会是一场势均力敌的比赛,而且我不会对深蓝获胜感到太惊讶。

While the news of the potential draw created a crisis in the Kasparov camp, among the Deep Blue team members, the main topic before game three was whether Garry would continue to play anti-computer chess. Our first feeling was that he would start to play normal chess because his anti-computer chess tactics were not doing too well. He did win game one, but Deep Blue could have secured a draw close to the end of the game. From our point of view, Garry was a little lucky that Deep Blue did not mix it up further and possibly even win the game. Game two was an outright disaster for him. He could have secured a draw as well but, like Deep Blue in game one, he did not see the drawing line either. Anyhow, the seed of Garry’s game two loss was clearly his desire to play anti-computer chess, which led him to reach awful positions that were supposedly good against other computers, but turned out to be dismal against Deep Blue. But this was an analysis based on pure chess merit. Our amateur psychological analysis of Garry’s mental state led to a very different conclusion. He had a very strong will; most people would even say that he could be stubborn at times. He came to the match with a set plan, and that apparently was to play anti-computer chess. It was most likely that he would not change his plan unless it proved to be a complete failure. A good question to ask, now that we are looking back, is what would have happened if he had played normal chess for the entire rematch? I don’t know the answer, but I do believe that it would have been a close match as well, and I would not be too surprised by a Deep Blue match win.

当第三场比赛开始时,整个深蓝团队都相信加里无论如何都会坚持反电脑象棋。所以当他打1.d3时,我们并不感到惊讶。乔是这个游戏的第一位操作员。1.d3 步可能从未在最高水平的国际象棋中下过。讽刺的是,这是我们在 Deep Thought 时代用来对抗另一个国际象棋程序的一步棋。观众席上的解说员对于这一举动,感到相当震惊。迈克·瓦尔沃:“哦,天啊。” 莫里斯·阿什利:“谨慎的举动。这场比赛,震撼人心。这场比赛拥有一切。” Yasser Seirawan:“我认为我们有一个新的开局动作。”

By the time game three started, the entire Deep Blue team believed that Garry would stick with anti-computer chess no matter what. So when he played 1. d3, it came as no surprise to us. Joe was the first operator for this game. The move 1. d3 probably had never been played at the highest level of chess. Ironically, it was a move that we had used against another chess program back in Deep Thought days. The commentators in the auditorium were quite shocked by the move. Mike Valvo: “Oh, my God.” Maurice Ashley: “A cagey move. A shock of shocks in this match. This match has everything.” Yasser Seirawan: “I think we have a new opening move.”

在最初的几个动作中,加里似乎试图欺骗深蓝去打西西里防守的白队,但是,他有一个额外的动作,这对于西西里的​​一些防线来说可能是至关重要的。深蓝的 4.…d6,而不是有风险的 d5,回避了这种可能性。

During the first few moves, Garry appeared to be trying to trick Deep Blue into playing the White side of the Sicilian defense, but, with him having an extra move, which could be critical in some of the Sicilian lines. Deep Blue’s 4.… d6, instead of the risky d5, sidestepped this possibility.

4.…d6 之前的位置。

The position before 4.… d6.

在加里的 5.NC3 之后,比赛又回到了已知的英语开局位置。在这个位置上,英国大师迈克尔·亚当斯(Michael Adams)曾扮演黑方Be7。由于 Michael 玩了这一场游戏,深蓝被赋予了玩 5....Be7 的强烈倾向。许多评论人士批评此举有点被动。这场比赛之后,穆雷更改了用于计算已知位置移动偏差的代码。Michael 是一位非常优秀的国际象棋棋手,但仅仅因为 Be7 下过一次棋就对他有很大的偏见,似乎有些过分了。

After Garry’s 5. Nc3, the game had transposed back into a known English opening position. In this position, British Grandmaster Michael Adams once played Be7 as Black. Deep Blue was given a hefty bias to play 5.… Be7 as a result of this one single game played by Michael. Many commentators criticized the move as a little bit passive. After this game, Murray changed the code used to compute the move bias in known positions. Michael was a very good chess player, but giving Be7 a hefty bias just because he had played it once, seemed excessive.

第 14 步之后的位置。

The position after move 14.

深蓝的14.…c5之后,Garry稍稍占据优势,但根据赛后有时间考虑位置的大师们表示,深蓝的位置似乎很稳固。一些大师的第一印象是这个位置显然对加里有利。英国大师约翰·纳恩 (John Nunn) 评论道:“确实,深蓝以一种在人眼看来非常奇怪的方式演奏开局,因为它不符合任何英语开局的正常演奏模式。理想情况下,怀特希望让他的王翼棋子移动,但是实现这一目标并不容易。” 有时,每秒搜索两亿个位置会产生不寻常的新解决方案。

After Deep Blue’s 14.… c5, Garry had a slight edge but, according to the Grandmasters who had had time to think about the position after the game, Deep Blue’s position appeared to be solid. Some Grandmasters had the first impression that the position clearly favored Garry. The British Grandmaster John Nunn commented, “True, Deep Blue has played the opening in a way which appears very odd to human eyes, because it doesn’t fit any of the normal patterns for playing with the English opening. Ideally, White would like to set his kingside pawn in motion, but it isn’t easy to achieve this.” Sometimes, searching two-hundred million positions per second produces unusual new solutions.

第 22 步之后的位置。

The position after move 22.

比赛的第一个关键时刻是深蓝打出22....Qa5,Garry回答23.Bd2。此时深蓝有两个选择,打Qxa3或者Nxg4。机器一开始更喜欢玩 Nxg4,但显然 Garry 玩得太多,所以它玩了 Qxa3。这一举动赢得了棋子,但作为交换,加里获得了丰厚的补偿。

The first critical moment of the game came after Deep Blue played 22.… Qa5, and Garry replied 23. Bd2. At this point, Deep Blue had two choices, to play Qxa3 or Nxg4. The machine at first preferred to play Nxg4, but apparently Garry would get too much play, so it played Qxa3. This move wins a pawn, but in exchange, Garry obtained serious compensation.

比赛的第二个关键时刻到来了,深蓝打出 26.…Bh7,让加里在比赛剩下的大部分时间里都无法阻挡其光方主教。我们对这一举动决定并不太满意,对深蓝如何评估这一立场对其有利也不满意。充其量,地位大致相等。本场比赛后,主教被禁止参加比赛的处罚增加了。另一种走法 Bh5 似乎是可以玩的,但这将是一场不同的游戏,双方都有机会。我们增加惩罚的主要目的是鼓励深蓝在有机会的时候尝试激活主教。

The second critical moment of the game came when Deep Blue played 26.… Bh7, allowing Garry to shut out its light square bishop for most of the rest of the game. We were not too happy with this move decision, nor were we happy about how Deep Blue was assessing the position as being in its favor. At best, the position was about equal. The penalty for bishops being shut out from the game was increased after this game. The alternative move, Bh5, appeared to be playable, but that would have been a different game with chances for both sides. Our main purpose in increasing the penalty was to encourage Deep Blue to try to activate the bishop whenever the opportunity presented itself.

比赛的其余部分并没有太引人注目。深蓝的棋步 40.... Bc7,还兵,是最后一个困难的棋步。我在游戏进行到一半时取代了乔作为操作员。加里在这场比赛的最后几招中不太高兴。深的蓝色的位置稍微差一点,但它已经建立了一个牢不可破的堡垒。最后五招,深蓝只是来来回回地洗牌,加里也没有什么建设性的办法。他知道这是一场平局,我知道这是一场平局,世界其他地方也知道这是一场平局,但在第二场比赛发生之后,加里无法让自己立即提出平局。如果赢了怎么办?当他在绝对、肯定地确定并提出抽签之前深思熟虑了很长时间时,我为他感到遗憾。

The rest of the game was not too remarkable. Deep Blue’s move 40.… Bc7, returning the pawn, was the last difficult move. I had replaced Joe as the operator in the middle of the game. Garry was not too happy during the last few moves in this game. Deep Blue was in the slightly worse position, but it had set up an unbreakable fortress. For the last five moves, Deep Blue just shuffled its king back and forth, and there was nothing constructive that Garry could do. He knew it was a draw, I knew it was a draw, and the rest of the world knew it was a draw but after what happened in game two, Garry could not bring himself to offer a draw immediately. What if there was a win? I felt sorry for him when he deliberated long and hard before he was absolutely, positively sure and offered the draw.

26 之前的位置... Bh7。

The position before 26.… Bh7.

40 之前的位置... Bc7。

The position before 40.… Bc7.

这场比赛结束后我非常兴奋。这是我们第一次在严肃的比赛中打成堡垒平局。当我们到达礼堂时,我还没准备好面对加里的爆发。

I was quite excited after this game. It was the first time we had ever played into a fortress draw in a serious game. I was not ready for Garry’s outburst when we got down to the auditorium.

他首先抱怨深蓝在他们所有的比赛中花费了最长的时间,当时它在第二局拒绝了Qb6这手棋。我们谁都不确定加里在说什么,但无论如何,他的记忆是错误的。深蓝在第二场比赛中两次拒绝了Qb6,当时它得到了36.axb5和37.Be4。它在第 36.axb5 棋上花费了大约 6.5 分钟,这比在 axb5 之前的棋 35.Bxd6 上花费的大约 15 分钟要短得多。此外,我几乎可以肯定,在之前对阵加里的比赛中,深蓝在其他动作上花费了超过十五分钟的时间。深蓝第二次拒绝Qb6,只花了三分钟多一点,这是它的正常分配时间。目前尚不清楚 Garry 指的是 36.axb5 还是 37.Be4。他在比赛中的一些言论似乎表明他更关心37.Be4,然而这只是一个走位动作,当然并没有放慢节奏。赛后,几位计算机国际象棋程序员表示,他们的程序将下 37.Be4,但评估函数略有变化。加里的下一句话似乎表明36.axb5让他感到惊讶。

He first complained that Deep Blue spent the longest amount of time in all their games when it rejected the move Qb6 in game two. None of us were sure what Garry was saying but in any case, his recollection was wrong. Deep Blue rejected Qb6 twice in game two, at the time when it made 36. axb5 and 37. Be4. It spent about 6.5 minutes at move 36. axb5, which was far shorter than the fifteen minutes or so it spent on 35. Bxd6, the move right before axb5. Furthermore, I am almost certain that Deep Blue had spent more than fifteen minutes on other moves in earlier games against Garry. The second time Deep Blue rejected Qb6, it spent just over three minutes, which was its regular allocated time. It was not clear whether Garry was referring to 36. axb5 or 37. Be4. Some of his remarks during the match seemed to indicate that he was more concerned about 37. Be4, which, however, was just a positional move and was certainly not played at a slower pace. After the match, several computer chess programmers indicated that their programs would play 37. Be4 with slight changes in their evaluation function. Garry’s next remark seemed to indicate that 36. axb5 surprised him.

他接着说道:“我牺牲了两个棋子,我还有一些机会。我认为这些机会还不够。” 我当时不知道他在说什么。现在回想起来,这一切都是有道理的。他在礼堂里见过亚西尔所说的两兵牺牲线。那条线不太管用。当时,我们没有人意识到深蓝的产品线是一个完全不同的牺牲。加里随后描述说,他在计算机上尝试了这些线,但没有人看到黑棋在两兵牺牲线中挽救游戏的任何方法。

He then said, “I sacrificed two pawns and I had some chances. I don’t think that these chances are enough.” I had no idea what he was saying at the time. Now, looking back, it all made sense. He had seen the two-pawn sacrifice line that Yasser was talking about in the auditorium. That line did not quite work. At the time, none of us realized that the line Deep Blue had was an entirely different sacrifice. Garry then described that he tried out the lines on his computers, but none of them saw any way for Black to save the game in the two-pawn sacrifice line.

他接着说:“从计算机的角度来看,这台机器没有达到基本的平局。” 我一听这话,心里自言自语道:“嗯?” 第二场比赛结束时的潜在平局对于计算机来说并不是一个基本的平局。要么他得到了关于计算机功能的非常糟糕的建议,要么他真的不了解国际象棋计算机。两种解释都很难相信。他确实有优秀的电脑国际象棋人员与他一起工作,如果他试图让弗里茨找到平局,他就会知道平局远非易事。如果深蓝错过了抽签,那么抽签显然不可能是初级抽签,除非我们有错误。加里的话完全没有道理。

He followed, “This machine missed, from a computer point of view, an elementary draw.” When I heard this, I said to myself, “Huh?” The potential draw at the end of game two was hardly an elementary draw for a computer. Either he was getting very bad advice about what computers could do, or he genuinely did not understand chess computers. Both explanations were hard to believe. He did have good computer chess people working with him, and had he tried to get Fritz to find the draw, he would have known that the draw was far from easy. If Deep Blue missed the draw, the draw obviously could not be an elementary one, unless we had a bug. Garry’s remark just did not make sense.

乔尔接过麦克风,简单地解释了发生的事情。我认为他用外行人的话很好地解释了一切。

Joel was given the microphone and explained what had happened in a straightforward way. I thought that he did a good job explaining everything in laymen’s terms.

然而,加里继续抱怨深蓝在拒绝 Qb6 上花了多少时间(只花了 6.5 分钟),抽签是多么简单,计算机不应该错过基本抽签,却还能够拒绝 Qb6 的举动。他还想看到动作的变化。

Garry, however, continued to complain about how much time Deep Blue had spent on rejecting Qb6 (it spent only 6.5 minutes), how simple the draw was, and how the computer should not have been able to miss the elementary draw, and yet have been able to reject the Qb6 move. He also wanted to see move variations.

这时,莫里斯·阿什利说道:“卡斯帕罗夫,如果我没听错的话,或者也许我说错了。你认为这场比赛中可能存在某种人为干预吗?”

At this point, Maurice Ashley said, “If I am reading you correctly, Kasparov, or maybe I’m speaking out of turn. Do you think that there may have been some kind of human intervention during this game?”

加里说:“这让我想起1986年马拉多纳对阵英格兰时打进的那粒著名进球,他说那是上帝之手。” 马拉多纳用手将足球击入球网,裁判误判进球有效。在我看来,加里显然是在指责我们作弊,尽管非常间接。我们每个人都气愤不已。我很想抓住麦克风,走到舞台上的深蓝模型面前,说:“好吧,我们打开这个盒子,看看鲍比(费舍尔)是否在里面?” 多年来,加里对我来说几乎就像一个朋友。他尽可能接近成为朋友,但又不是朋友。深蓝战队和卡斯帕罗夫阵营的比赛对双方来说都非常好。突然,我觉得我不想再和他有任何关系了。

Garry said, “It reminds me of the famous goal which Maradona scored against England in 1986. He said it was the Hand of God.” Maradona knocked the soccer ball into the net with his hand and the referee mistakenly allowed the goal to count. It seemed to me that Garry was apparently accusing us of cheating, although very obliquely. Every one of us was fuming. I was tempted to grab the microphone, go to the Deep Blue model on the stage, and say, “Well, shall we open up this box and see whether Bobby (Fischer) is inside?” Over the years, Garry had been almost like a friend to me. He was as close as possible to being a friend without being one. The competition between the Deep Blue team and the Kasparov camp had been very good for both sides. Suddenly, I felt that I no longer wanted to have anything to do with him.

乔尔再也受不了了,“比赛结束后,我们很乐意准确地分享我们在这起案件中看到的情况。但我认为,加里给弗里茨或任何其他计算机一个位置并说:“这是计算机行为,这就是深蓝一定在想或者深蓝会做的,这绝对是一个错误。” 我认为他从与“深蓝”对战的比赛中看出,这不是普通的计算机,并且与他见过的任何计算机的水平完全不同。所以也许他应该正视这样一个事实:深蓝可以做很多他认为不可能的事情。”

Joel could not take it any more, “After the match, we’ll be happy to share exactly what it saw in this case. But I think that it’s definitely a mistake for Garry to give a position to Fritz or any other computer and say, ‘This is computer behavior and this is what Deep Blue must be thinking or what Deep Blue would do.’ I think he’s seen from the games that he’s played against Deep Blue that it is no ordinary computer and plays at an entirely different level from any computer that he’s seen. So maybe he should come to grips with the fact that Deep Blue can do a lot of things that he did not think were possible.”

乔尔本可以更圆滑一些。加里从不比赛结束后,即使在深蓝向他提供了有关两步走法的日志后,他还是开始了交锋。我认为部分问题在于他要求提供错误动作的日志。他应该要求提供第二局中第 35 步 Bxd6 步的日志。如果他看到了 35.Bxd6 日志,他就会意识到深蓝在完成对 Bxd6 的 10 层搜索时已经看到了三兵牺牲。

Joel could have been a little bit more diplomatic. Garry never did come to grips, not even after he was given Deep Blue’s log for the two moves in question, after the match. I think part of the problem was that he asked for the log of the wrong moves. He should have asked for the log of move 35. Bxd6 in game two. Had he seen the 35. Bxd6 log, he would have realized that Deep Blue had already seen the three-pawn sacrifice when it finished its 10-ply search for Bxd6.

加里的另一个问题确实是他不知道深蓝的能力。在第 36 步中避免 Qb6 的决定主要是战术性的,并且主要涉及计算。如果他真的研究过 Deep Thought 或 Deep Thought II 游戏,他就会意识到 Deep Blue 的前辈们已经能够产生远远超出其他计算机能力的战术,即使其他计算机已经扩展到 Deep Thought 或 Deep Thought II 的水平。搜索速度。Deep Blue 的速度比 Deep Thought 或 Deep Thought II 快了近千倍。

The other part of Garry’s problem was indeed that he did not know what Deep Blue was capable of. The decision to avoid Qb6 on move 36 was mostly tactical, and involved mainly calculations. If he had really studied Deep Thought or Deep Thought II games, he would have realized that Deep Blue’s predecessors were already capable of producing tactics way beyond the capabilities of other computers, even if the others had been scaled up to Deep Thought or Deep Thought II’s search speed. And Deep Blue was close to a thousand times faster than either Deep Thought or Deep Thought II.

最后但并非最不重要的一点是,三兵的牺牲是相当困难的,至少对于其他计算机来说是这样,而且对于加里来说似乎也是如此。法国国际象棋记者皮埃尔·诺洛特 (Pierre Nolot) 曾收集过 11 个国际象棋棋局,他表示任何计算机都无法在合理的时间内解决这些棋局。Deep Thought II 在锦标赛期间解决了其中的几个问题,一天之内可能解决了八个或九个问题。Deep Thought II 解决的一些局面可能仍然超出了最好的商业国际象棋程序8 的范围,即使给了整整一年的时间。36.Qb6 之后可能出现的三兵牺牲可能属于同一类复杂性。

Last, but not least, the three-pawn sacrifice is quite hard, at least for other computers, and, it seems, for Garry as well. Pierre Nolot, a French chess journalist, once collected eleven chess positions that he said were unsolvable by any computer in reasonable time. Deep Thought II solved several of them in tournament time, and maybe eight or nine of them within a day. Some of those positions that Deep Thought II solved are probably still beyond the best commercial chess programs8, even when given a full year of time. The three-pawn sacrifice that could arise after 36. Qb6 might be in the same class of complexity.

死里逃生

A Narrow Escape

在第四场比赛开始之前,人们对加里是否会出现存在严重怀疑。他的经纪人欧文·威廉姆斯(Owen Williams)在电话里待了一会儿,当得知加里正在路上时,每个人都松了口气。

Before the start of game four, there was some serious doubt as to whether Garry would show up at all. His manager, Owen Williams, was on the phone for a while, and everyone was relieved when the report came that Garry was on his way.

直到几个月后,深蓝团队对发生的事情几乎一无所知。以下是对相关人士的采访整理。

The Deep Blue team was mostly in the dark about what was happening until months later. The following is a reconstruction from interviewing the people involved.

上诉委员会成员肯·汤普森 (Ken Thompson) 表示,第二场比赛后的晚上,他收到了加里的请求。他要他检查一下游戏的两份日志,确认一切都是光明正大的。肯实时看到了日志,很清楚发生了什么事,但他想在向加里汇报之前再看一遍以恢复记忆。

According to Ken Thompson, who was on the appeals committee, the night after game two he received a request from Garry. He wanted him to go over the game two log and verify that everything was above board. Ken had seen the log in real time and was well aware of what happened, but wanted to see it again to refresh his memory before he reported back to Garry.

Ken随后要求CJ打印休息日的日志,CJ欣然同意了这一要求。肯通知加里,请求已被接受。后来,CJ答应在第二天中午左右给Ken提供两个有问题的游戏登录信息。肯显然告诉上诉委员会的其他成员,问题已得到解决。根据 C J 的回忆,上诉委员会的某个人(可能是 Monty Newborn)随后告诉他问题已经解决。不幸的是,这导致 Ken 没有在指定时间获取日志,因为 CJ 认为问题不再重要。大约在这个时候,卡斯帕罗夫阵营中的某人告诉 CJ,加里想亲眼看看日志。随后欧文·威廉姆斯表示,IBM 已同意让加里在比赛期间访问所有比赛日志,但后来又违背了承诺。据我所知,IBM 的任何人都没有同意他可以在比赛进行期间访问比赛日志。一种可能性是卡斯帕罗夫阵营的人可能误解了肯告诉他们的话。

Ken then asked C J for a printout of the log on the rest day and C J readily agreed to the request. Ken informed Garry that the request had been accepted. Later, C J promised to give Ken the game two log in question around noon the next day. Ken apparently told the other members of the appeals committee that the problem was taken care of. According to C J’s recollection, someone from the appeals committee, probably Monty Newborn, then told him that the problem was resolved. Unfortunately, this led to Ken not getting the log at the specified time, as C J thought that the problem no longer mattered. Somewhere around this time, someone in the Kasparov camp told C J that Garry wanted to see the log for himself. Owen Williams was later quoted as saying IBM had agreed to let Garry have access to all the game logs during the match but then reneged on the promise. As far as I know, no one from IBM had ever agreed that he could have access to the game logs while the match was still going on. One possibility was that people in the Kasparov camp might have misinterpreted what Ken told them.

如果由深蓝团队决定,我们绝不会同意让加里在比赛期间访问游戏日志。深蓝的游戏日志包含其“内心想法”,包括游戏将如何进行、假设游戏中没有出现的走法它会做什么、它认为加里的最佳走法应该是什么等等。如果他有日志,他就会有一个完整的路线图,说明深蓝在各种情况下的行为,无论是正确的还是错误的。即使是新手,只要有游戏日志、良好的记忆力,再加上优秀大师团队的帮助,只需记住获胜台词,就能始终如一地击败深蓝。加里除了可能是有史以来最强的大师之外,还拥有惊人的记忆力。允许他直接访问深蓝的游戏日志,就相当于给了他一个银盘上的深蓝硅头,并附带了 70 万美元。

If it were up to the Deep Blue team, we would never have agreed to give Garry access to the game logs during the match. Deep Blue’s game logs contained its “inner thoughts” on how the games would proceed, what it would have done given hypothetical moves that did not show up in the games, what it thought Garry’s best moves should be, and so on. If he had the logs, he would have a complete road map on how Deep Blue would behave, rightly or wrongly, in various circumstances. Even a rank beginner, armed with the game logs, a very good memory, and the help of a good team of Grandmasters, would be able to beat Deep Blue consistently by merely memorizing the winning lines. Garry, besides being possibly the strongest Grandmaster ever, has a phenomenal memory. Granting him direct access to Deep Blue’s game logs would be equivalent to giving him Deep Blue’s silicon head on a silver platter, with $700,000 on the side.

不管怎样,当肯在第三场比赛前几分钟向CJ询问日志时,CJ可能以为他是在问加里是否会得到日志,并给出了一个干脆的“不”。肯然后告诉加里说这个请求被拒绝了。在发现第二场比赛的最终位置可能是平局后,加里很可能已经不在最佳状态了。事实上,他在第三场比赛中有白棋,而那场比赛以平局结束,这一事实可能也没有帮助他的情绪。难怪赛后他对深蓝队进行了猛烈的抨击。这一系列的事件并没有让他的爆发在我看来完全可以原谅,但确实让它变得更容易理解。

Anyway, when Ken asked C J minutes before game three about the log, C J probably thought that he was asking whether Garry would be given the log, and gave a flat “no”. Ken then told Garry that the request was turned down. Garry was most likely already not in the best frame of mind after discovering that the final position in game two could be a draw. The fact that he had White in game three and that game ended a draw probably did not help his mood either. No wonder he lashed out at the Deep Blue team after the game. This whole sequence of events did not make his outburst completely excusable in my mind, but did make it more understandable.

到第四场比赛的那天早上,Ken 开始对 CJ 生气,并表示除非打印出来,否则他不会来操作室。IBM 安全人员给当时在新泽西州办公室的 Ken 打了电话,表示等他回来后就可以打印出来。与此同时,欧文表示加里不会参加比赛。加里后来说他从未说过他不会参加比赛。整个事情一直悬而未决,直到接近比赛时间,据报道加里正在路上。肯终于拿到了日志,并在第四场比赛前十分钟通知加里收到了日志。

By the morning of the day for game four, Ken was starting to get angry with C J, and said that he would not come to the operations room unless the printout was produced. A phone call was placed by IBM security to Ken, who was in his office in New Jersey, that the printout would be available for him when he came back. Meanwhile, Owen was saying that Garry would not play. Garry later said that he never said he would not play. The whole thing was up in the air until close to game time, when it was reported that Garry was on his way. Ken finally got the log and informed Garry about its receipt ten minutes before game four.

现在,回到游戏本身。

Now, back to the game itself.

在第四场比赛中,亚西尔发表了这样的评论:“在我看来,深蓝正处于深渊之中”,这让礼堂里充满了笑声。第四局给了我整场比赛最焦虑的时刻,尽管深蓝确实打平了。这也是一场充满丰富性的游戏。加里做出了深刻的棋子牺牲,深蓝很难完全理解。后来我们还发现了第一场比赛中出现的国王安全评估bug。我确信,如果我更深入地研究这个游戏,我会发现许多新的想法来改进深蓝。

During game four, Yasser made the comment, “It looks to me like Deep Blue is in deep doo-doo”, which filled the auditorium with laughter. Game four gave me the most anxious moment of the match, even though Deep Blue did manage to draw. It was also a game full of richness. Garry played a profound pawn sacrifice that Deep Blue had a hard time fully appreciating. Afterwards we also found the king safety evaluation bug that had shown up in game one. I am certain that if I look deeper into the game, I will find many new ideas to improve Deep Blue.

加里这次用 1.…c6 回答了深蓝的 1.e4,卡罗·坎恩的防守。他再一次没有发挥他最喜欢的西西里防守,尽管现在这已不再令人惊讶。我们满心期待他继续玩奇怪的东西。然而,加里的下一步行动,2.…d6,让礼堂和新闻发布室的评论员摸不着头脑,“我们该怎么称呼这个东西?”

Garry answered Deep Blue’s 1. e4 with 1.… c6 this time, the Caro Kann defense. Once again, he did not play his pet Sicilian defense, although by now this was no longer a surprise. We were fully expecting him to continue to play strange stuff. Garry’s next move, 2.… d6, however, had the commentators in the auditorium and the pressroom scratching their heads, “What do we call this thing?”

米格尔·伊列斯卡斯大师在这场比赛的操作室里。早期我也在手术室里。米格尔是我见过的最开朗的人之一。对于深蓝的举动,他几乎总能找到一些好话来评价。当我们需要振作起来时,他是一个很棒的人,在这场比赛中我们也需要振作起来。

Grandmaster Miguel Illescas was in the operations room for this game. I was also in the operations room for the early part. Miguel is one of the most cheerful people that I have ever met. He almost always found some good things to say about Deep Blue’s moves. He was a great person to have around when we needed cheering up, and in this game we needed cheering up.

2.…d6 之后的位置。

The position after 2.… d6.

米格尔很早就对深蓝的机会相当乐观。到了第二十步,机器似乎已经解决了所有的开局问题,并准备好开始一些真正的行动。就在此时,加里释放了惊人的典当牺牲20.... e5。米格尔看到这个举动,本能地称这是一个很好的牺牲。深蓝不会采取这一行动,但认为这种牺牲是合理的。

Miguel was quite optimistic about Deep Blue’s chances early on. By move twenty, the machine seemed to have all the opening problems solved and was ready to start some real action. Right at this point, Garry unleashed an amazing pawn sacrifice 20.… e5. When Miguel saw this move, he instinctively called it a good sacrifice. Deep Blue would not play the move but assessed the sacrifice as sound.

卡斯帕罗夫 20 之前的位置... e5!.

The position before Kasparov’s 20.… e5!.

接下来,深蓝又打出了一些在人眼看来相当奇怪的招式。26. b5受到了不同评论者的批评和赞扬。加里本人对此举给予了高度评价。我们不喜欢这一举动,并在第二天引发了评估软件错误搜寻。乔最终发现并修复了与此动作以及第一场游戏中相应动作相关的错误。深蓝在下b5之前的奇怪举动似乎是由相同的错误引起的。

Next, Deep Blue played some moves that were quite strange to human eyes. 26. b5 was both panned and praised by different commentators. Garry himself spoke highly of the move. We did not like the move, and it triggered an evaluation software bug hunt the next day. Joe was the one who finally found and fixed the bug relating to this move and the corresponding move from game one. Deep Blue’s strange moves right before playing b5 appeared to be caused by the same bug.

深蓝26之前的位置... b5。

The position before Deep Blue’s 26.… b5.

当深蓝奇怪的b5之后局势开始变得清晰时,米格尔开始重新考虑。局势持续恶化,就连米格尔也开始担心起来。深蓝开始评估这个位置对加里来说更好。这一天对我们来说是光明的,三场比赛有两场白人比赛,但现在情况看起来很糟糕。

As the position started to clarify after Deep Blue’s strange b5, Miguel was having second thoughts. The position continued to deteriorate, and even Miguel was getting worried. Deep Blue was starting to evaluate the position as better for Garry. The day had started brightly for us, with two Whites for three games, but now things were looking bad.

就在这个时候,手术室里来了一位惊喜的客人。IBM 董事长兼首席执行官郭士纳 (Lou Gerstner) 决定拜访该团队。《纽约时报》记者布鲁斯·韦伯恰好在房间里。第二天,卢对这场比赛的戏剧性评论发表在《纽约时报》上:“我只是认为我们应该把这看作是世界上最伟大的国际象棋棋手和……之间的国际象棋比赛”,短暂的停顿,“加里·卡斯帕罗夫。

At this juncture, we had a surprise guest in the operations room. IBM Chairman and CEO, Lou Gerstner, had decided to pay the team a visit. New York Times reporter Bruce Weber happened to be in the room. Lou’s dramatic comment on the match was printed in the New York Times the next day: “I just think we should look at this as a chess match between the world’s greatest chess player and …”, a short pause, “Garry Kasparov.”

此后,我们谁都不愿意破坏这一刻并告诉卢深蓝可能遇到了麻烦。卢停留了大约五分钟,然后就离开了。

After this, none of us was willing to spoil the moment and tell Lou that Deep Blue might be in trouble. Lou stayed for about five minutes and then left.

当米格尔有时间再次查看棋盘位置时,他变得相当悲观,是我在整场比赛中见过的最悲观的人。他认为深蓝将无法幸免于难,这个结局看起来很不利。然而,深蓝对局面的评估从未低于半兵。默里当时正在运营深蓝公司。如果米格尔对这个职位如此不满,穆雷的感觉一定更糟。

When Miguel had time to look at the board position again, he became quite pessimistic, the most pessimistic that I ever saw him in the entire match. He thought that Deep Blue would not be able to survive the ending, which looked unfavorable. Deep Blue’s assessment of the position, however, never got below half a pawn down. Murray was operating Deep Blue at the time. If Miguel was so unhappy about the position, Murray must have been feeling even worse.

在加里第 43 步时,深蓝自我终结。比赛前,Joe 添加了一段代码来监视并行搜索的效率,如果效率下降得太低,则终止程序本身9。乔的想法是,在这种情况下,重新开始会更好。此实例是整个复赛期间执行自我终止代码的两次之一。在第三局中,深蓝也自我终结,但那是在比赛的早期,就在深蓝第八步之后,而加里仍在思考他的下一步。我们必须等到他采取行动后才能在那场比赛中我们自己的时间重新启动程序。

On Garry’s move forty-three, Deep Blue self-terminated. Before the match, Joe added in a piece of code that monitored the efficiency of the parallel search and terminated the program itself if the efficiency dropped too low9. Joe’s idea was that a fresh start would be preferable in this case. This instance was one of the two times during the entire rematch that the self-termination code got executed. In game three, Deep Blue also self-terminated, but it was very early in the game, right after Deep Blue’s move eight, and while Garry was still thinking about his move. We had to wait until he made his move before restarting the program on our own time in that game.

深蓝第四局自我终止后,我接任了接线员。加里简单地问发生了什么事,当我告诉他程序必须重新启动时,他很满意。加里的四十三手恰好给了深蓝一次合法的答复,所以他清楚地知道机器回来后会下什么棋。我们的问题给了他自由思考的时间,同时,深蓝的时钟也在滴答作响。复赛后,我们听说加里对深蓝多次“崩溃”感到不安,并且他可能暗示这些“崩溃”会分散注意力和可疑。就我个人而言,我更希望 Joe 不使用像自我终止代码这样的机制,但我不站在 Joe 的立场上。然而,我强烈怀疑加里真的因为两次自我终止而分心,当然没有什么值得怀疑的,尤其是在第四场比赛的情况下。你不能违反国际象棋规则并走非法棋。

I took over as the operator after Deep Blue’s game four self-termination. Garry asked briefly what happened and was satisfied when I told him that the program had to be restarted. Garry’s move forty-three happened to give Deep Blue only one legal reply, so he knew precisely what the machine would play when it came back. Our problem gave him free time to think on the position, meanwhile, Deep Blue’s clock was ticking. After the rematch, we had heard that Garry was upset that Deep Blue “crashed” numerous times and that he may have suggested that the “crashes” were distracting and suspicious. Personally, I would have preferred that Joe did not use a mechanism like the self-termination code, but I was not in Joe’s shoes. However, I strongly doubted that Garry was really distracted by the two instances of self-termination and there was certainly nothing to be suspicious about, especially in the case of game four. You cannot violate the chess rules and play an illegal move.

当我离开操作室时,我预计会发生一场灾难,但当深蓝重新启动时,情况看起来好多了。之前我只是听米格尔说没希望了。现在我正在自己检查深蓝的台词,那些出现的台词似乎表明是平局。棋盘上的位置是车结局。国际象棋芯片包含残局ROM(只读存储器),其中封装了许多相关的车结局知识。深蓝还可以在线访问五件车结局数据库,尽管在这款游戏中数据库并不重要。深蓝第五十六手之后,局面已经达到了明显的平局。加里张开双手,没有动手,说道:“抽牌?” 这个抽签提议是非正统的。通常情况下,人们在采取行动后立即提出平局。我立即接受了这个提议。

I was expecting a disaster when I left the operations room, but by the time Deep Blue was restarted, the position looked much better. Previously, I was just hearing Miguel saying that it was hopeless. Now I was examining Deep Blue’s lines for myself, and those coming up seemed to point to a draw. The position on the board was a rook ending. The chess chip contains endgame ROMs (Read Only Memory) that encapsulate a lot of the relevant rook ending knowledge. Deep Blue also had access to the five-piece rook ending database on line, although in this game the database did not matter. After Deep Blue’s move fifty-six, the position had reached a clear draw. Garry opened up his hands and said, before making any move, “Draw?” This draw offer was unorthodox. Normally, one offers a draw right after making a move. I immediately accepted the offer.

加里在残局阶段有赢过吗?目前还不清楚。许多分析师声称他们找到了获胜路线,但经过仔细检查,它们都不起作用。

Did Garry have a win somewhere in the endgame phase? It is unclear. Many analysts claimed that they found winning lines, but on closer examination none of them works.

四局比赛过后,双方各积两分。

The match was all tied up with two points each after four games.

喝一两杯后

After a Drink or Two

第四场比赛之后是两个休息日。与费城的比赛中,第四场和第五场之间只有一个休息日,但在复赛的谈判中,加里明确要求额外休息一天。仅场地和设备的租赁费,IBM 就多花了一天至少 100,000 美元。

Game four was followed by two rest days. In the Philadelphia match, there was only one rest day between game four and game five but during the negotiation for the rematch, Garry explicitly requested the extra rest day. The extra day probably cost IBM at least $100,000 more, in leases alone, for the venue and the equipment.

当然,对于深蓝团队来说,休息日并不是真正的“休息”日。我们三个人和大师们一起努力解决比赛中出现的问题,或者在练习赛或测试位置中发现潜在的问题。乔修复了关键评估错误。其中一场比赛促使穆雷对评估函数进行了概念上的改变。我们向加里本人学习。我对 Deep Blue Jr 进行了大量的测试,发现了一些错误,乔立即修复了这些错误。大师们致力于开局,尽管现在很明显深蓝的开局准备工作可能不会受到测试。当我们完成错误修复后,大师们与 Deep Blue Jr 进行了一些测试游戏,以确保修改后的程序确实有所改进。这持续了两个休息日的大部分时间。

Of course, rest days were not really “rest” days for the Deep Blue team. The three of us, along with the Grandmasters, were hard at work trying to resolve problems that showed up in the match games, or finding potential problems in practice games or test positions. The critical evaluation bugs got fixed up by Joe. One of the match games triggered Murray to make a conceptual change to the evaluation function. We were learning from Garry himself. I ran Deep Blue Jr through a vast set of test positions and located a few bugs, which Joe promptly fixed. The Grandmasters worked on the openings, although it was clear by now that Deep Blue’s opening preparations probably would not be tested. When we completed the bug fixes, the Grandmasters played a few test games against Deep Blue Jr to make sure that the modified program was indeed an improvement. This went on for most of the two rest days.

在第五场比赛中,加里恢复了他在第一场比赛中的基本设置,但略有偏差。第四步,深蓝决定用主教换取马,从而获得主教对的优势。但换来主教对,深蓝会有更好的棋子发展。在操作室里,我们可以从深蓝的主线上看到这一点。放弃主教是一个完全合理的选择。深蓝团队中没有人对这一举动考虑太多。这只是一个举动。但我们也知道加里可能对此有问题。卡斯帕罗夫阵营的一些评论表明,他们对深蓝评估函数的看法非常简单。第一局比赛结束后,卡斯帕罗夫阵营(亚西尔·塞拉万也是如此)显然认为深蓝对主教的重视程度远高于骑士,并且会采取严重的扭曲措施来保留主教。加里在赛后新闻发布会上的言论表明,他对于深蓝用主教换骑士的决定确实感到惊讶。我想知道他是否考虑过“深蓝”可能确实比他所认为的更加复杂的可能性。

In game five, Garry reverted to the basic setup he played in game one, but with a slight deviation. On move four, Deep Blue decided to trade off its bishop for his knight, giving him the advantage of the bishop pair. But in exchange for the bishop pair, Deep Blue would have better piece development. In the operations room, we could see this from Deep Blue’s main line. Giving up the bishop was a perfectly reasonable choice. None of us on the Deep Blue team thought much about the move. It was just a move. But we also knew that Garry might have problems with it. Some comments from Kasparov’s camp indicated that they had a very simplistic view of Deep Blue’s evaluation function. After game one, the Kasparov camp (as did Yasser Seirawan) apparently thought Deep Blue valued the bishop much higher than the knight, and it would go through serious contortions to preserve it. Garry’s remark at the post-match press conference indicated that he was indeed surprised by Deep Blue’s decision to trade off the bishop for the knight. I wondered whether he considered the possibility that Deep Blue might, indeed, be far more sophisticated than he gave it credit for.

一些媒体成员显然已经接受了加里关于深蓝如何评估立场的简单化观点。这场比赛期间,我去新闻发布室接受了几次采访,令我惊讶的是,反复有人问我们是否已经调整了。赛前主教和马的相对值。记者的本意是好的,但问题却像是在问:“你刚才是用左手拿起书的。你改变了你的左手和右手的相对值吗?”

Several members of the press apparently had taken up Garry’s simplistic view about how Deep Blue evaluated a position. During this game, I went to the pressroom to give a few interviews, and was surprised by repeated questions about whether we had adjusted the relative values of bishop and knight before the game. The reporters meant well, but the questions felt like being asked, “You used your left hand to pick up the book just now. Did you change the relative values of your left hand and your right hand?”

11.…h5 之前的位置。

The position before 11.… h5.

加里第十一步棋后,深蓝在棋子发展上取得了显着的领先优势。此时,深蓝出手 h5,将其 h 列棋子推入两个方格。加里开始在记分表上记下这一动作,然后他看了一眼,看了一眼正在为深蓝操作的穆雷。穆雷对加里微微点头,表示深蓝确实出手了。加里很惊讶。我们手术室里的所有人也是如此。乔尔的第一反应是:“我不喜欢这个举动。但我也不能说这有什么问题。” 在礼堂里,莫里斯·阿什利的第一反应是“不!” 亚西尔:“加里是一个快乐的露营者!谁在给这台机器编程?” 莫里斯:“我的意思是我们已经看到了一些奇怪的举动,我们已经看到了一些奇怪的举动。这真是一个奇怪的举动!” 米格尔一开始显然也不太喜欢这个举动,但想了想,他说道:“好吧,也许前十名的选手中会有一两个人会玩这个,但前提是喝一两杯酒之后。” 我从来没有问过米格尔他的意思。但我是这样解释他的评论的。h5这招是在向Garry发出信号,“如果你城王这边,我就攻击你。” 正常情况下,就算是前十的高手,也不敢如此咄咄逼人。深蓝显然不知道它正在扮演加里·卡斯帕罗夫。加里赛后评论表示,他对这一举动感到非常震惊,“没有电脑玩h5!” 很有趣的是,比赛过程中,新闻发布室的反应是h5是电脑棋,有些人在“电脑”前插入“愚蠢”。比赛的分析师后来称赞此举是一个复杂的位置理念。

After Garry’s eleventh move, Deep Blue had a significant lead in piece development. At this point, Deep Blue produced the move h5, pushing its h-file pawn two squares. Garry started to jot down the move on his score sheet, then he took a double take and glanced at Murray, who was operating for Deep Blue. Murray gave Garry a slight nod, indicating that Deep Blue did play the move. Garry was surprised. So were all of us in the operations room. Joel’s first reaction was, “I don’t like the move. But I can’t say what is wrong with it.” In the auditorium, the immediate response from Maurice Ashley was “No!” Yasser: “Garry is a happy camper! Who’s been programming this machine?” Maurice: “I mean we’ve seen some strange moves and we’ve seen some strange moves. That’s a strange move!” Miguel apparently did not quite like the move at first either but after a moment’s thought, he said, “Well, maybe one or two of the top ten players would play it, but only after a drink or two.” I never did ask Miguel what he meant. But I was interpreting his comment the following way. The move h5 was signaling to Garry, “If you castle king side, I will attack you.” Normally, even a top-ten player would not dare to be so aggressive. Deep Blue obviously had no idea that it was playing Garry Kasparov. Garry’s post match comment indicated that he was quite shocked by the move, “No computer plays h5!” It was quite interesting that, during the game, the reaction in the pressroom was that h5 was a computer move, with some people inserting “stupid” before “computer”. Analysts of the game later praised the move as a sophisticated positional idea.

两年后,h5这一举动成为网络上争议的焦点。当我第一次听到加里赛后评论“没有电脑玩h5”时,我不知道加里是在赞扬我们还是在指责我们。比赛结束后,克拉拉·卡斯帕洛娃问儿子是否有什么问题。他回答说这是一个典型的电脑游戏,所以他肯定有可能是在表扬我们。然而,1999年,当他接受国际象棋记者采访时,他的言论似乎暗示着一种指控。国际象棋记者在互联网上发布了我们所理解的加里的断言,即没有计算机下 h5,因此我们作弊了。几个小时内,网上有人指出,至少有一个商业象棋程序也玩h5。这位国际象棋记者在采访期间碰巧知道,加里在几英尺外的自己的笔记本电脑上也有同样的商业国际象棋程序。所以他不仅错误地指责了我们,而且似乎他甚至懒得去检查他拥有的国际象棋程序是否下的是 h5。不知道为什么商业棋程序下的是h5。这可能是好的编程,也可能是糟糕的编程,或者只是运气不好。

Two years later, the move h5 was at the center of a controversy over the Internet. When I first heard Garry’s post-match comment that “no computer plays h5”, I had no idea whether Garry was praising or accusing us. Right after the game itself, Clara Kasparova asked her son whether there was anything amiss. He replied that it was a typical computer game so there was certainly the possibility that he was praising us. In 1999, however, when interviewed by a chess journalist, he seemed to imply that his statement was an accusation. The chess journalist posted on the Internet what we understood to be Garry’s assertion that no computer played h5, and therefore we had cheated. Within hours, someone on the Internet pointed out that at least one commercial chess program also played h5. The chess journalist happened to know that Garry had the very same commercial chess program on his own laptop a few feet away during the interview. So he not only wrongly accused us, but it appears he did not even bother to check whether any of the chess programs he owned played h5. I don’t know why the commercial chess program played h5. It could be good programming, bad programming or just dumb luck.

当我看到深蓝的 h5 举措时,我确切地知道是什么硬件评估功能促使了这一举措。在复赛前的最后两个月的芯片设计中,我对King安全评估的硬件进行了大幅改动。在国王城堡之前,硬件会计算三组国王安全评估,一组用于王侧易位,一组用于后侧易位,一组用于留在中心。真正的王者安全评价是三者的加权线性组合,权重基于三者的相对排名以及易位的难度。正常的国际象棋程序有一个简单得多的国王安全评估。为了避免国王安全问题,国际象棋程序员通常会积极阻止他们的程序下 h5 之类的棋步,因为它会损害王侧易位的潜在国王安全性,尽管实际上,程序可能能够安全地城堡后侧。由于深蓝可以判断它是否可以自行安全地在后侧建立城堡,因此我们就允许它做任何它想做的事。在对局位置上,深蓝总能安全地占据后翼,因此从它的角度来看,h5 棋是完全可以下的。

When I saw the move h5 from Deep Blue, I knew precisely what hardware evaluation features prompted the move. During the last two months of chip design before the rematch, I added drastic changes to the hardware for king safety evaluation. Before the king castles, the hardware computes three sets of king safety evaluations, one for kingside castling, one for queenside castling, and one for staying in the center. The real king safety evaluation is the weighted linear combination of the three, with the weighting based on the relative ranking of the three, and difficulty of making the castling moves. Normal chess programs have a far simpler king safety evaluation. To avoid problems with king safety, chess programmers usually actively discourage their programs from playing moves like h5, because it damages the potential king safety for castling kingside, even though in reality, the programs might be able to castle queenside safely. Since Deep Blue could tell whether it could castle queenside safely on its own, we just allowed it to do whatever it wanted. In the game position, Deep Blue could always castle queenside safely, and therefore move h5 was perfectly playable from its point of view.

深蓝第二十二招之后,评估是稍占优势。我变得乐观起来,“我们能赢得这场比赛吗?” 正当我兴奋的时候,我被叫到新闻发布室接受采访。加里显然认为他也遇到了麻烦。然而,当我回到手术室时,他已经巧妙地避开了任何潜在的问题。大约在这个时候,深蓝的一手棋导致了一些位置问题,但加里并没有发挥出最好的延续性。

After Deep Blue’s move twenty-two, it assessed that it had a slight edge. I was getting optimistic, “Could we possibly win this game?” Just when I was getting excited, I was called up to the pressroom to give interviews. Garry apparently thought he was in trouble as well. By the time I got back to the operations room, however, he had skillfully tiptoed around any potential problems. Around this time, Deep Blue played a move that led to some positional problems but Garry did not play the best continuation.

到了第四十步,深蓝评估形势为平局。加里后来表示,此时他看到了比赛将如何在九步之后结束。观众席上的解说员直到最后才看到比赛结束。比赛的结局相当漂亮。乍一看,加里的棋子似乎很难阻止。深蓝,而不是试图阻止过去的棋子,威胁要将死。在加里抵御了交配威胁后,深蓝只是将其国王向前推进,忽略了经过的棋子。最后的游戏位置让加里准备将他传递的棋子提升为皇后,但这一切都无济于事。深蓝即将启动基于重复检查的绘图序列。

By move forty, Deep Blue assessed the position as being even. Garry later stated that he saw at this point how the game would end in nine more moves. The commentators on the auditorium stage did not see the game finish until the very end. The game finish was quite pretty. At first glance, Garry had a passed pawn that seemed very difficult to stop. Deep Blue, instead of trying to stop the passed pawn, threatened to checkmate. After Garry fended off the mating threat, Deep Blue just marched its king forward, ignoring the passed pawn. The final game position had Garry ready to promote his passed pawn to a queen, but it was all to no avail. Deep Blue was about to initiate a drawing sequence based on repetition checks.

第 5 局的最终位置。白方下棋。

The final position in game 5. White to move.

平局确定后,加里留在棋桌旁,开始用俄语与母亲交谈。很快游戏室就挤满了人。加里的几个人以及肯和 CJ 都在场。在手术室里,我们对发生的事情感到茫然。我们被告知暂时不要去游戏室。很快就有消息称,加里希望将所有比赛日志打印、密封,并交给比赛仲裁员卡罗尔·贾雷茨基 (Carol Jarecki)。前一天,他正式要求将所有游戏日志打印出来并放入密封的信封中。上诉委员会投票反对要求IBM 这样做,但也建议 IBM 应该考虑这样做。加里被告知委员会的决定。就在比赛之前,IBM 内部做出了一项决定,我们将尊重他的请求,推翻了之前的决定。所以问题很快就得到了解决,尽管也有一些小问题。电脑打印机卡纸,需要一段时间才能修复,而且没有足够的纸张,也没有足够的时间来打印所有游戏日志。仅有的第五场比赛的日志已打印。第二天,两套完整的比赛日志被打印出来并交给了卡罗尔·贾雷茨基和肯·​​汤普森。

Garry stayed at the chess table after the draw was agreed, and began talking to his mother in Russian. Soon the game room was filled with people. Several of Garry’s people were there as well as Ken and C J. In the operations room, we were at a loss as to what was happening. We were told not to go over to the game room yet. Word soon came that Garry wanted all the game logs to be printed, sealed, and placed in the possession of the match arbiter, Carol Jarecki. The day before, he had made a formal request that all the game logs be printed and placed in sealed envelopes. The appeals committee had voted against requiring IBM to do so, but had also suggested that IBM should consider doing it. Garry was told about the decision of the committee. Just prior to the game, an internal IBM decision was made that we would honor his request, reversing an earlier decision. So the problem was quickly resolved, although not without some hiccups. The computer printer had a paper jam that took a while to fix, and there was not enough paper nor enough time to print all the game logs. Only the log for game five was printed. The next day, two complete sets of game logs were printed and given to Carol Jarecki and Ken Thompson.

卡罗尔将第五场比赛的七十二页日志安全地放入公文包后,她前往新闻发布室向媒体简要介绍了所发生的情况。卡罗尔还曾担任 1995 年卡斯帕罗夫与阿南德世界冠军赛的比赛仲裁员。卡罗尔在结束简报时对两场比赛进行了比较,“我告诉你一件事。这场比赛比世锦赛精彩多了。”

After Carol got the seventy-two pages of game five’s log safely in her briefcase, she went up to the pressroom and briefed the press about what happened. Carol had the distinction of also being the match arbiter for the 1995 Kasparov vs. Anand World Championship match. Concluding her briefing, Carol made a comparison between the two matches, “I’ll tell you one thing. This match is a lot more exciting than the World Championship.”

她指的不仅仅是游戏。

She did not mean just the games.

30万美元的赌博

A $300,000 Gamble

五场比赛过后,双方战成2.5分。深蓝上一场棋是白棋。除非加里能够以某种方式赢得黑方的最后一场比赛,否则历史将会被创造。深蓝要么是第一台在常规比赛中追平世界国际象棋冠军的计算机,要么更好的是,第一台在常规比赛中击败世界国际象棋冠军的计算机。

After five games, the match was tied at 2.5 points each. Deep Blue had White in the last game. Unless Garry could somehow manage to win the last game from the Black side, history would be made. Deep Blue would be either the first computer to tie the World Chess Champion in a regulation match, or better still, the first computer to beat the World Chess Champion in a regulation match.

黄牛们全力出击,将 25 美元的门票转售至 500 美元。如果您购买一张 500 美元的门票,那么您将为每分钟比赛时间支付约 10 美元,因为第六场比赛持续不到一个小时。

The scalpers were out in full force, reselling the $25 tickets for as much as $500. If you had bought one of the $500 tickets, you would have paid about $10 for every minute of game time, as game six lasted less than an hour.

就我个人而言,在第六场比赛之前,我感到一种平静。默里、乔和我经过长期努力才达到这个阶段。比赛打成平手。深蓝有白棋,所以我们至少平局的机会还是很大的。然而,我们无法真正控制接下来会发生什么。“该怎样就怎样,”我告诉自己。是的,我十二年的努力也许终于结束了,但我只能眼睁睁地看着。我唯一想要的就是双方打一场精彩的比赛,最好是精彩的比赛,历史学家会这样评价这场比赛:“双方都不应该输。” 我不在乎出现的是反电脑加里还是普通加里。我只是想要一场精彩的比赛,以适应在如此激烈的比赛之后可能具有历史意义的时刻。

Personally, before game six, I felt a sense of calm. Murray, Joe, and I had worked long and hard to reach this stage. The match was tied. Deep Blue had White, so our chance of at least drawing the match was quite good. We had, however, no real control over what would happen next. “What will be will be,” I told myself. Yes, my twelve-year endeavor might finally be over, but I could only watch at this point. The only thing I wanted was just a good—preferably great—game by both sides, and that the historians would say about the match, “Neither side deserved to lose.” I did not care whether the anti-computer Garry or the normal Garry showed up. I just wanted a great fight as befitting the potentially historic occasion after such a closely contested match.

除了公关人员之外,大多数 IBM 员工也都精神抖擞。如果深蓝打平或赢得比赛,他们将面临一场风暴。当比赛仍打平时,加里已经很不满意了。如果他人生中第一次输掉比赛,他会说什么?是的,你没有看错。加里有在1997年复赛之前,他的职业生涯从未输过一场国际象棋比赛。在他们的第一场世界冠军赛中,他差点输给阿纳托利·卡尔波夫,但由于阿纳托利的健康状况似乎不佳,那场比赛被取消了。有些人担心加里会对输掉比赛做出愤怒的反应。我对这个想法感到惊讶。复赛在很多方面甚至比世界国际象棋锦标赛的比赛还要大。公关人员正在为火山喷发做准备。IBM 团队被要求盛装出席,并特别要求在闭幕式上不要微笑,特别是如果深蓝赢得了比赛。公关人员不想在风暴爆发时火上浇油。

Most of the IBMers were also in great spirits, with the exception of the PR people. They were bracing for a firestorm should Deep Blue draw or win the match. Garry had been dissatisfied when the match was still tied. What might he say if he lost a match for the first time in his life? Yes, you read it right. Garry had never lost a single chess match in his professional life before the 1997 rematch. He had come close to losing to Anatoly Karpov in their first world title match, but that match was cancelled when Anatoly’s health seemed to be failing. Some were concerned that Garry would react angrily to losing a match. I was surprised at this notion. The rematch was in many ways bigger than even the World Chess Championship matches. The PR people were making preparations for a volcanic eruption. The IBM team was asked to dress up for the occasion, and also asked specifically not to smile during the closing ceremony, especially if Deep Blue won the match. The PR people did not want to fan the fire should a firestorm erupt.

比赛开始前,礼堂内气氛热烈。电视节目《大卫·莱特曼深夜秀》《杰·雷诺今夜秀》中有关复赛的视频片段被播放,赢得了观众的笑声和掌声。

The atmosphere in the auditorium before the start of the game was festive. Video clips from the TV shows ‘Late Night with David Letterman’ and ‘The Tonight Show with Jay Leno’ about the rematch were shown, to laughter and applause from the audience.

乔是上一场比赛的操作员,穆雷是替补。我已经明确地决定十二年后的今天我不会工作。这将是我作为旁观者的一天。

Joe was the operator for the last game, with Murray as the backup. I had decided positively and definitively that I would not work on this particular day after twelve years. This was going to be my day as just a spectator.

比赛开始的方式与第四场相同,由卡罗·坎恩防守。这次加里没有像第四场比赛那样打奇怪的边线,而是打主线。深蓝也为白棋打了主线;事实上,这是加里饰演怀特时用来对抗阿纳托利的主线。加里身处熟悉的领域,尽管是在棋盘的另一边。

The game began the same way as game four, with a Caro Kann defense. Instead of playing a weird sideline as he did in Game four, this time Garry played the main line. Deep Blue also played the main line for White; in fact, the main line that Garry as White used against Anatoly. Garry was in a familiar territory, albeit from the opposite side of the chessboard.

在加里的第七步棋上,花了近两分钟后,他下了 7.... h6。深蓝立刻回复了Nxe6,放弃了一个马换了一个棋子。加里表现得有些惊讶,但接下来的几招下得很快,深蓝也是如此。双方仍处于预定状态。

On Garry’s seventh move, after spending nearly two minutes, he played 7.… h6. Deep Blue instantly replied Nxe6, giving up a knight for a pawn. Garry acted a little bit surprised, but then played the next few moves very fast, as did Deep Blue. Both sides were still in book.

在礼堂里,亚西尔说加里犯了错误,改变了棋法。他还说加里处于恐惧和痛苦之中。我们手术室里的人,包括肯,都不相信这一点。令人难以置信的是,具有传奇记忆力的加里竟然忘记了他已经多次演奏过的开场白,尽管通常是白队的。加里的表情也更多的是惊讶,而不是恐惧和痛苦。那些情绪会在以后出现。

In the auditorium, Yasser was saying that Garry had blundered and transposed the move. He also said that Garry was in terror and distress. None of us in the operations room, including Ken, believed this. It was inconceivable that Garry, with his legendary memory, could have forgotten an opening line that he had played many times, albeit usually from the White side. Garry’s expression was also more one of surprise rather than terror and distress. Those emotions would come later.

那么到底发生了什么?第二天,报纸援引卡斯帕罗夫阵营的一位国际大师的话说,7.... h6此举是早些时候商定的。加里本人在几个月后的一次采访中表示,他对玩 7.... h6 的决定感到遗憾。因此,这一举动绝非偶然,正如赛后新闻报道所描述的那样。这是故意设计的。

So what did happen? An International Master in Kasparov’s camp was quoted the next day in the newspaper that the 7.… h6 move was one agreed upon earlier. Garry himself stated months later in an interview that he regretted the decision to play 7.… h6. So the move was never a slip of the finger as it was characterized in news articles immediately following the match. It was played by design.

卡斯帕罗夫的 7 之后的位置... h6。

The position after Kasparov’s 7.… h6.

那么加里为什么要采取这一行动呢?在马牺牲之后,黑棋的局面充其量被认为是困难的。在人类玩家玩的游戏中是这样,但在计算机和人类玩家之间玩的游戏中却不是这样。商业国际象棋程序显然在避免白棋输掉比赛方面存在严重问题!几位大师曾尝试与顶级商业国际象棋程序对弈黑方,并赢得了每一场比赛。这一事实直到比赛结束后才被曝光,但那时媒体已经转向其他报道了。加里在第六局第七步棋中所下的棋是非常危险的反电脑棋步。

So why did Garry play the move? Black’s position was generally considered difficult at best after the knight sacrifice. This is true in games played by human players, but it is not true in games played between computers and human players. The commercial chess programs apparently had serious problems avoiding losing the game as White! Several Grandmasters had tried playing the Black side of the game positions against the top commercial chess programs and were able to win every single game. This fact did not come out until well after the match, but by then the media had moved on to other stories. What Garry played in game six on move seven was a very risky anti-computer chess move.

加里是怎么想到玩 7.... h6 的?在赛后新闻发布会上,他似乎将责任归咎于他的助手。还有其他可能的来源。当时的几个顶级商业国际象棋程序在其开本中都明确禁止下深蓝下的骑士牺牲。因此,显然,许多商业国际象棋程序员都知道他们的程序无法走牺牲路线。曾有一些传闻在比赛中,Garry 认为深蓝正在使用其中一个商业国际象棋程序的开局书,因此可以想象,他期望深蓝会像其他程序一样撤退其马。这会给加里一个比他在正常路线上所能达到的更好的位置。在深蓝确实扮演了骑士牺牲这一“不太可能”的事件中,加里的前副手之一、大师蒂莫先科(Timoshchenko)在 1997 年 3 月号的《国际计算机象棋协会杂志》上发表了一篇精彩的文章蒂莫申科(Timoshchenko)与弗里茨(Fritz)下的完全相同的路线,弗里茨是加里的计算机国际象棋顾问弗雷德里克·弗里德尔(Frederic Friedel)分发的国际象棋程序。季莫先科给出了判断:“马牺牲后,黑方有足够的防守可能性。”

How did Garry get the idea to play 7.… h6? In the post-match news conference, he seemed to blame his helpers. There were other possible sources. Several of the top commercial chess programs at the time were explicitly prohibited in their opening books from playing the knight sacrifice that Deep Blue played. So, apparently, lots of the commercial chess programmers knew that their programs could not play the sacrificial line. There were some rumors during the match that Garry thought Deep Blue was using the opening book of one of the commercial chess programs, so it was conceivable that he was expecting Deep Blue to retreat its knight as the other programs did. This would have given Garry a much better position than he could have achieved in the normal lines. In the “unlikely” event that Deep Blue did play the knight sacrifice, well, there was this nice article written by Grandmaster Timoshchenko, one of Garry’s former seconds, in the March 1997 issue of the International Computer Chess Association Journal. Timoshchenko played exactly the same line against none other than Fritz, the chess program distributed by Garry’s computer chess consultant, Frederic Friedel. Timoshchenko gave the verdict, “After the knight sacrifice, Black has enough possibilities for defense.”

所以有很多潜在的理由选择7.... h6。首先,深蓝可能会撤退骑士。在这种情况下,加里很容易扳平比分。其次,深蓝可能不“理解”如何攻击。那么,加里就赢得了比赛,成为人类万众瞩目的对象。(虽然在我看来,归根结底,深蓝可能是更重要的人类成就。)第三,这是一个延伸,如果深蓝确实“理解”了这个立场并击败了他,旁观者总是可以说最后一场比赛不算在内,而且考虑到第二场比赛可能会平局(就像第一场比赛一样),他是这场比赛的实际获胜者。不管你信不信,确实有人用这个论点来暗示深蓝是这场比赛的失败者。如果认为第三种情况是加里策划的,那就太愤世嫉俗了。

So there were lots of potential reasons to select the move 7.… h6. First, Deep Blue might just retreat the knight. In this case, Garry would have equalized very easily. Second, Deep Blue might not “understand” how to attack. Then, Garry would have won the game and become the toast of the human race. (Although Deep Blue, in my opinion, might be the more important human achievement when all was said and done.) Third—and this one a stretch—if Deep Blue did “understand” the position and beat him, a bystander could always say that the last game did not count and, given that game two could have been drawn (as game one could have), he was the actual winner of the match. Believe it or not, there were actually people who used this argument to suggest that Deep Blue was the loser of the match. It would be way too cynical to believe that Garry had planned this third scenario.

深蓝在第 11 步 Bf4 棋时出局。它对该位置的第一个评估是它接近一个棋子上升,尽管它实际上在棋盘上的材料上下降了大约两个棋子。深蓝看到了接近三个棋子的位置补偿。大约在这个时候,米格尔突然出现,询问深蓝对这个职位的评价如何。我们告诉他后,他笑得很开心。我们所有人都知道比赛实际上已经结束了。深蓝一旦掌握了主动权,并且知道自己掌握了主动权,就没有办法阻止。棋盘上的位置实际上已经在大约一个月前出现在我们的实验室中,当时其他大师正在研究深蓝的开篇书。当时,大师们也懒得去扮演这个角色;他们只是看了深蓝对这个职位的评价,然后就开始做其他的工作。他们都曾在接收端处于类似位置。讨论时米格尔不在美国,所以他还没有看到深蓝对该职位的评价。也许黑棋可以抽出位置,但可能没有人类玩家可以对抗深蓝。看来加里的 30 万美元赌博会给他带来适得其反的效果。

Deep Blue was out of its opening book on move 11. Bf4. Its first evaluation of the position was that it was close to a pawn up, even though it was really down about two pawns in material over the board. Deep Blue saw close to three pawns worth of positional compensation. At about this time, Miguel popped in and asked what was Deep Blue’s evaluation of the position. After we told him, he smiled broadly. All of us knew that the match was effectively over. Once Deep Blue had the initiative, and knew that it had the initiative, there was no stopping it. The position on the playing board had actually shown up in our lab about a month earlier when the other Grandmasters were working on Deep Blue’s opening book. At the time, the Grandmasters did not bother to play out the position either; they just looked at Deep Blue’s evaluation of the position and moved on to other work. They all had been on the receiving end in similar positions. Miguel was not in the United States at the time of the discussion, so he had not yet seen Deep Blue’s evaluation of the position. Maybe Black could draw the position, but probably no human player could hold the position against Deep Blue. It appeared that Garry’s $300,000 gamble would backfire on him.

深蓝11之后的位置……Bf4。

The position after Deep Blue’s 11.… Bf4.

加里在棋盘上的表情很快就从惊讶变成了苦恼,而深蓝则拒绝了快速获得的小材料并继续进攻。比赛临近结束时,加里开始用俄语与克拉拉交谈。在比赛仍在进行时与观众交谈是违反国际象棋规则的。目前尚不清楚他对母亲说了些什么。也许他告诉克拉拉,再玩下去已经没有意义了。Joe为深蓝打完19.c4后,Garry用询问的眼神看着Joe。乔微微点头,证实了自己的猜测。深蓝赢得了主要材料。加里随后立即辞职。在接下来的几天里,他辞职的视频片段将在世界各地多次播放。辞职后,加里与乔握手,然后双臂高举,径直走出房间。这是一个超现实的场景。比赛持续了不到一个小时。

Garry’s demeanor at the board quickly changed from surprised to distressed, as Deep Blue spurned a quick gain of small material and continued its attack. Near the end of the game, Garry started talking to Clara in Russian. It is against the rules of chess to talk to spectators while a game is still going on. It was not clear what he said to his mother. Perhaps he told Clara that there was no sense in playing on any more. After Joe played 19. c4 for Deep Blue, Garry looked at Joe in a questioning manner. Joe nodded slightly, confirming what he was suspecting. Deep Blue was winning major material. Garry then promptly resigned. The video segment of his resignation would be played many times, around the world, in the next few days. After his resignation, Garry shook hands with Joe and just walked straight out of the room with both arms up in the air the entire time he was walking out. It was a surreal scene. The game had lasted less than an hour.

在Garry辞职的那一刻,我突然感到很累。十二年的工作终于结束了。我应该心里很高兴,但内心却感到空虚。这场比赛感觉太简单了,尽管事后看来并非如此。如果没有我们前一年的努力,加里可能会赢得比赛。我的一部分也感到被抢劫了。我不是国际象棋棋手,但我内心的棋手绝对感到失望。无论输赢,我都希望最后一场比赛是一场真正的战斗。我希望这场胜利能成为另一场像第二场比赛一样伟大的比赛,除了最后的错误。如果这必须是另一场失败,我希望它是另一场像第一场比赛一样的艰苦战斗的失败,最好没有深蓝的糟糕动作。

At the moment of Garry’s resignation, I suddenly felt very tired. The twelve years of work were finally over. I should have been exulting, but I was feeling empty inside. The game felt too easy, although in hindsight it wasn’t. Without our hard work the year before, Garry might have won the game. A part of me also felt robbed. I am not a chess player, but the chess player in me was definitely disappointed. Win or lose, I had wanted the last game to be a real fight. I wanted the win to be another great game like game two, except without the final mistake. If it had to be another loss, I wanted it to be another hard-fought loss like game one, preferably without the bad moves from Deep Blue.

因此,在克劳德·香农 (Claude Shannon) 提出如何编写计算机下棋的建议近五十年后,对圣杯的追求终于结束了。它并没有以有史以来最伟大的国际象棋比赛结束,但比赛如何结束并不取决于我们。

So, almost fifty years after Claude Shannon made the proposal on how to program a computer to play chess, the quest for the holy grail was finally over. It did not end with the greatest chess game ever, but how the match ended was not up to us.

比赛最终名次。

The final position of the match.

    令人好奇的是,那场比赛的开场白是后弃兵开局的塔拉什变体。

    For the curious, the opening line played in that game was the Tarrasch variation of the Queen’s Gambit opening.

    要玩好某些开局,可能需要程序能够识别国际象棋位置评估功能,而其他国际象棋程序(包括深蓝)无法识别这些功能。在这种情况下,我们就需要添加新的硬件,或者至少是新的软件,以便机器能够很好地发挥空缺。

    To play certain openings well may require a program to recognize chess positional evaluation features that no other chess programs, including Deep Blue, recognize. In that case, we would need to add new hardware, or at least new software, so that the machine can play the openings well.

    深蓝发现了一些新的台词,其中包括被大师拒绝的台词。在接受后弃兵开局中,它找到了它喜欢的白棋牺牲品(e4)。大师们认为深蓝此举过于雄心勃勃,因为典当的补偿似乎微不足道。几个月后的复赛,同样的招式在宗师级别上发挥,并且取得了不错的效果。

    Deep Blue found some of the new lines, including one that was rejected by the Grandmasters. In the Queen’s Gambit Accepted opening, it found a pawn sacrifice (e4) for White that it liked. The Grandmasters thought that Deep Blue’s move was too ambitious as the compensations for the pawn seemed to be tenuous. A few months after the rematch, the same move was played at the Grandmaster level, and with good results.

    一份报告称,卡斯帕罗夫团队在分析完为什么 44.... Rf5 不好时也笑了。

    One report stated that the Kasparov team also burst out laughing when they finished their analysis on why 44.… Rf5 was no good.

    在撰写本文时,您可以通过链接到 Deep Blue 比赛站点,然后从网站 www.chess.ibm.com 链接到各个游戏来找到游戏日志。IBM 可能不会无限期地维持该站点。

    At the time of this writing, you can find the game logs by following the links to the Deep Blue match site and then to the individual games from the web site www.chess.ibm.com. It is possible that IBM may not keep the site up indefinitely.

    对于国际象棋爱好者来说,三兵牺牲线的开头如下: 36. Qb6!? Qe7!(替代的两兵牺牲走法 Rd8 被 37. Be4 驳斥!) 37. ab5 Rab8。现在白方的最佳路线是 38。Qe3!ab5 39. Be4,抛弃三兵牺牲,达到与游戏类似的位置。如果白方决定接受三兵牺牲,那么我们有 38.Qa6 e4!39. Be4 (39. Rel Qe5 40. Re4 Qh2 41. Kf2 Bg3 42. Kf3 Bd6 平局) Qe5 40. Bf3 Rd8 和黑方将获得显着补偿。例如,如果 41.Qa3, Re8 42.Re2 Qh2 43.Kf1 Re2 44.Be2 Re8 45.Bh5 Qf4 46.Bf3 Re3 47.Qb2 Qh2 抽牌。如果 41. Ra3,Qe3 42. Kf1 Qd3 43. Be2 Qf5 44. Kgl Qe5 45. Kfl (45. Bf3 Re8) Re8 46. Bf3 Bc7!47. d6 Bb6 48. d7 Red8 49. Rel Qd6,黑方可能获胜。

    For the chess enthusiasts, the three-pawn sacrifice line starts as follows: 36. Qb6!? Qe7! (the alternative two-pawn sacrifice move Rd8 is refuted by 37. Be4!) 37. ab5 Rab8. And now the best line for White is 38. Qe3! ab5 39. Be4, spurning the three-pawn sacrifice to reach a position similar to the game. If White decides to accept the three-pawn sacrifice, then we have 38. Qa6 e4! 39. Be4 (39. Rel Qe5 40. Re4 Qh2 41. Kf2 Bg3 42. Kf3 Bd6 draw) Qe5 40. Bf3 Rd8 and Black would get significant compensation. For example, if 41. Qa3, Re8 42. Re2 Qh2 43. Kf1 Re2 44. Be2 Re8 45. Bh5 Qf4 46. Bf3 Re3 47. Qb2 Qh2 draws. If 41. Ra3, Qe3 42. Kf1 Qd3 43. Be2 Qf5 44. Kgl Qe5 45. Kfl (45. Bf3 Re8) Re8 46. Bf3 Bc7! 47. d6 Bb6 48. d7 Red8 49. Rel Qd6, and Black might be winning.

    某些其他程序可能声称它看到了重复,这并非不可能,但如果确实如此,则很可能是由于使用换位表而导致的错误重复声明。

    It is not out of the question that some other program might claim that it sees the repetitions, but chances are that if it does, it would be due to false repetition claims resulting from the use of the transposition table.

    当您阅读本文时,情况可能会发生变化。在添加了单一扩展后,一些业余程序在 Nolot 位置上表现良好。一些商业程序员很可能也决定采用这个想法。

    This may change by the time you read this. A few amateur programs were doing well on the Nolot position after adding in singular extensions. Some commercial programmers may very well decide to adopt the idea as well.

    对于同时也是计算机科学家的读者来说,当并行系统出现死锁时,效率可能会下降得太低。此外,事情也可能因其他未知原因而出错。

    For the readers who are also computer scientists, the efficiency could drop too low when there is a deadlock in the parallel system. Also, things can go wrong for other unknown reasons.

 

 

结语

Epilogue


 

 

第13章

CHAPTER 13

下棋后的生活

Life after Chess


新闻发布会

The Press Conference

在赛后新闻发布会上,加里将这场比赛形容为“世界冠军输了”。然后他猛烈抨击 IBM,至少对我来说,他似乎在暗示第二场比赛需要人工干预。他还再次错误地将第二场比赛中重复检查的潜在平局描述为“简单”。还有一些虚张声势。“我个人向你保证,我会毫无疑问地将它撕成碎片。” “这不会有帮助,因为我们知道机器是如何玩的。” “我认为这台机器有太多弱点。” 更令人惊讶的说法之一是“但在我看来,它还没有准备好赢得一场重大比赛。” 我不知道他在想什么,但从所有媒体指标来看,复赛比他参加过的任何其他赛事都要大得多。仅复赛的“印象”数量,还不包括赛后宣传,是5到70亿。这可能是他任何一场世界锦标赛比赛的展示次数的 100 到 1000 倍。如果复赛不是一场大赛,我不知道什么才是大赛。

At the post-match press conference, Garry characterized the match as “lost by the world champion”. He then lashed out at IBM and he seemed, to me at least, to be hinting at human intervention in game two. He also once again wrongly characterized the potential draw by repetition checks in game two as “simple”. There was also some bravado. “I personally guarantee you, I’ll tear it to pieces without question.” “It will not help, because we know how the machines play.” “I think the machine has too many weaknesses.” One of the more surprising statements was “But it’s not yet ready, in my opinion, to win a big contest.” I don’t know what he had been thinking, but in all media metrics, the rematch was a far bigger event than any other that he had played in. The number of “impressions” registered for the rematch alone, not counting postmatch publicity, was 5 to 7 billion. This was probably 100 to 1000 times the number of impressions registered for any of his World Championship matches. If the rematch was not a big contest, I don’t know what is.

在第六场比赛中,加里说:“今天的比赛甚至不能算作比赛,因为它可能以前在其他地方发布过。当电脑对上e6时你就可以辞职了。我必须告诉你,我根本没有心情打球。” 听到这话,我在心里摇了摇头。当时,我对评论中完全缺乏体育精神感到震惊。我当时并不像现在那么确定加里刚刚进行了 30 万美元的赌博并输了。

On game six, Garry said, “Today’s game doesn’t even count as a game because probably it has been published before elsewhere. When the computer takes on e6 you can resign. I was not in the mood of playing at all, I have to tell you.” I shook my head in my mind when I heard this. At the time, I was just appalled at the total absence of sportsmanship in the comment. I was not as certain as I am now that Garry had just made a $300,000 gamble and lost.

回到第二场比赛,加里再次强调,“忘记今天的比赛吧。我的意思是,深蓝在五场比赛中没有赢过一场,因为在第二场比赛中,我在本可以强行放弃的时候辞职了。画。” 他很方便地忘记提及他没有看到绘图延续。按照同样的道理,我们也可以说他在这五局比赛中没有赢过一场,因为深蓝下了一个错误的棋子,输掉了第一局,就像他下了一个错误的棋子,认输,输掉了第二局一样。

Going back to game two, Garry again stressed, “Forget today’s game. I mean Deep Blue hasn’t won a single game out of the five because again in game two, I resigned when I could have forced a draw.” He conveniently forgot to mention that he did not see the drawing continuations. By the same argument, we could have also said that he did not win a single game out of the five, as Deep Blue played a wrong move to lose game one, just as he made a wrong move, resignation, to lose game two.

卡斯帕罗夫在第三场比赛后的新闻发布会上发表讲话。

Kasparov speaking in the press conference after game 3.

从左到右:迈克·沃尔沃、加里·卡斯帕罗夫、徐凤雄、CJ Tan、乔尔·本杰明、默里·坎贝尔、乔·霍恩和杰里·布罗迪。卡斯帕罗夫正要开始公开指控。

From left to right: Mike Volvo, Garry Kasparov, Feng-hsiung Hsu, C J Tan, Joel Benjamin, Murray Campbell, Joe Hoane, and Jerry Brody. Kasparov was about to begin his public accusations.

在未来的比赛中,加里希望,“有一个条件,IBM 作为球员,而不是赞助商。” 我不介意自己得到他所得到的皇家待遇。据我所知,IBM 的赞助商正在合理范围内满足他的每一个需求。比赛进行时,他并没有从IBM玩家那里得到游戏日志,但这只是为了获得不公平优势的请求。如果他可以自由进入,那场比赛就会是一场骗局。

On future matches, Garry wanted, “One condition, IBM as player, not a sponsor.” I would not mind for myself getting the royal treatment that he was getting. To the best of my knowledge, the part of IBM that was the sponsor was attending to his every need, within reason. He was not getting the game logs from the IBM player while the match was going on, but this was simply a request for an unfair advantage. The match would have been a sham if he were given free access.

紧接着又是一个大惊喜。我可能误解了卡斯帕罗夫的话,但看起来他是在指责深蓝团队在第一场比赛中作弊,也就是我们输掉的一场比赛:“你知道,它(反计算机策略)正在发挥作用,但突然停止了工作——突然深蓝找到了一种方法来打破典当链,并在一个非常非常方便的情况下开始对抗。”

This was followed by another big surprise. I may have misinterpreted what Kasparov was saying, but it appeared he was accusing the Deep Blue team of cheating in game one, the one game we lost: “You know, it [the anti-computer strategy] was working, but suddenly it stopped working—suddenly Deep Blue found a way just to break the pawn chains and start a confrontation in a very, very convenient situation.”

加里终于平静了一些:“不过,这台机器甚至能和世界冠军打出这样的水平,显然是一个历史性的成就。”

Garry eventually calmed down a little bit: “But obviously it’s a historical achievement that the machine was even able to play on such a level with the World Champion.”

更多关于未来比赛的信息:“我认为八场或十场比赛,一场正常的比赛。” 后来,他补充道:“我认为我们应该每隔一天比赛一次。你应该给一个人休息的时间。你知道,二十天,十场比赛,一场合适的比赛。”

More on future matches: “I think eight or ten games, a normal match.” Later, he added, “I think we should play every second day. You should give a human being time to rest. You know, twenty days, ten games, a proper match.”

就我个人而言,我有些怀疑能否找到他建议的比赛持续时间的赞助商。除去奖金之外,比赛延长的时间将使赞助商损失超过一百万美元。额外的日子可能会大大降低媒体的兴趣,尤其是一半只是休息日。NBA 锦标赛和职业棒球世界大赛各进行七场比赛是有充分理由的。媒体对一项体育赛事的全面关注很难超过一周。他的提议将导致一场持续近三周的活动,其中一半的时间没有发生任何事情。

Personally, I have some doubt whether a sponsor could be found for a match of the duration that he suggested. Excluding the prize fund, the extra length of the match would cost the sponsor in excess of one million dollars. The extra days could significantly reduce the media interest, especially with half being just rest days. There are good reasons why the NBA Championships and professional baseball’s World Series are seven games each. It is difficult to keep the full attention of the media on a sports event for much longer than a week. His proposal would lead to an event lasting nearly three weeks and with nothing happening on half of the days.

发布会持续了约三十分钟。整个过程中,深蓝团队都坐着听他的独白。我们被要求不要微笑来表达我们的喜悦,也不能表达我们对他的暗示的愤怒。我们刚刚完成了多年来的梦想和努力:我用了十二年,默里用了大约九年,乔用了七年。然而,我们不得不坐在那里听这些指控。这是我一生中最糟糕的经历之一,而本应是我最快乐的时刻。

The press conference lasted about thirty minutes. During the entire time, the Deep Blue team sat and listened to his monologue. We were asked not to smile to express our joy, nor could we voice our outrage at his implications. We had just accomplished what we had dreamt and worked for years: twelve years for me, about nine years for Murray and seven for Joe. And yet, we had to sit there and listen to these charges. It was one of the worst experiences of my life at a time that should have been my most joyous moment.

后来,乔说了一句至今我还记得的话:“你知道,他(加里)毁了这一切的乐趣。”

Afterwards, Joe said something that I remember to this day, “You know, he [Garry] ruined the joy of it.”

老帮派

The Old Gang

两个月后,我们迎来了更加欢乐的时刻。1997 年 7 月 29 日,在罗德岛州普罗维登斯举行的 AAAI 全国会议上,穆雷、乔和我获得了弗莱德金奖,表彰他们第一个建造或编程国际象棋机器并在一场比赛中击败世界象棋冠军的团队。价值 10 万美元的弗莱德金奖是复赛获胜者奖金的七分之一,除奖金受托人保留的额外一分钱外,均分为三份。

Two months later, we were at a more joyous occasion. On 29 July 1997, in Providence, Rhode Island, at the AAAI national conference, Murray, Joe and I received the Fredkin Prize for the first team to build or program a chess machine to defeat the World Chess Champion in a match. The $100,000 Fredkin Prize, which was one-seventh of the winner’s prize for the rematch, was split evenly three ways, excluding the extra cent that was kept by the trustee of the Prize.

复赛结束后,我们并不相信我们会获得弗雷金奖。复赛虽然满足了奖项的初衷,但不符合评奖委员会的要求比赛条件是几年前设定的,特别是比赛的长度和奖金的数额。Murray 和我加入 IBM 后不久,Fredkin 奖委员会主席 Berliner 博士以及我们在卡内基梅隆大学的前皇家反对派发表了一份公开声明,详细说明了领取该奖项所需的匹配条件。在发言的最后,柏林纳博士承认,他的比赛条件很可能不符合世界冠军和比赛赞助商的喜好。对于深蓝比赛,据我所知,在设置两场比赛时,IBM 甚至没有人费心去查看 Berliner 博士的比赛条件清单。

Right after the rematch, we did not believe that we would receive the Fredkin Prize. While satisfying the original intent of the Prize, the rematch did not conform to the Prize Committee’s match conditions set a few years back, in particular the length of the match and the size of the prize fund. Shortly after Murray and I joined IBM, Dr Berliner, the chairman of the Fredkin Prize Committee, and our former royal opposition at Carnegie Mellon, issued a public statement specifying the required match conditions for claiming the Prize. At the end of his statement, Dr Berliner admitted that it was perfectly possible that his match conditions might not be to the liking of the World Champion and the match sponsor. For the Deep Blue matches, as far as I know, no one at IBM even bothered to look at Dr Berliner’s list of match conditions when setting up the two matches.

弗雷德金奖委员会没有在复赛结束时发布任何公告;委员会似乎认为深蓝的比赛胜利不够充分。复赛的胜利对我们来说已经足够奖励了,所以我们没有追求获奖的可能性。令人惊喜的是,复赛大约一个月后,Raj Reddy 教授向 AAAI 通知我们,我们赢了。是什么改变了委员会的想法?我不知道。也许他们意识到,对于整个世界的所有意图和目的来说,探索已经结束了。

The Fredkin Prize Committee did not make any announcement at the conclusion of the rematch; it appeared that the committee considered Deep Blue’s match win inadequate. Winning the rematch was rewarding enough for us, so we did not pursue the possibility of claiming the Prize. It came as a pleasant surprise that, about a month after the rematch, Professor Raj Reddy informed us for AAAI that we had won it. What changed the mind of the committee? I have no idea. Maybe they realized that, for all intents and purposes to the world at large, the quest was over.

颁奖后,我向大会介绍了深蓝的贡献。在开始演讲之前,当我收到全场起立鼓掌时,我很感动。在纽约,第五场比赛(深蓝最好的比赛之一)结束后,游击队观众对我们发出了嘘声。这一次,我和朋友们在一起,他们明白这场比赛从来都不是真正的“人与机器”,而是“作为表演者的人与作为工具制造者的人”。无论比赛结果如何,当我们为获胜者欢呼时,我们是在为人类独一无二的成就而欢呼。1996年加里获胜时,深蓝队为他出色的个人表现欢呼雀跃。当深蓝在 1997 年获胜时,整个社会(不包括加里·卡斯帕罗夫)最终认识到了深蓝的本质,即人类创造的强大工具的进步。

After the awarding of the Prize, I gave a presentation to the assembly on the contribution of Deep Blue. I was moved when I received a standing ovation before I started the presentation. In New York, after game five, one of the better games by Deep Blue, the partisan crowd booed us. This time, I was among friends, friends who understood that the match was never really “man versus machine”, but rather “man as a performer versus man as a toolmaker”. Whatever the outcome of the match, when we cheered the winner, we were cheering for a unique human achievement. When Garry won in 1996, the Deep Blue team cheered and applauded his outstanding personal performance. When Deep Blue won in 1997, society at large, not including Garry Kasparov, finally recognized Deep Blue for what it was, namely, the advancement of a powerful tool created by human beings.

出席 AAAI 会议的还有我们 Deep Thought 时代的几位老朋友,他们因与 Deep Thought 的共同工作而获得艾伦·纽厄尔卓越研究奖。托马斯·阿南塔拉曼(Thomas Anantharaman)离开了华尔街公司,在纽约大学担任助理教授,从事与人类基因组项目相关的研究。安德烈亚斯·诺瓦齐克 (Andreas Nowatzyk) 离开 Sun Microsystems 并加入 Digital仪式结束后我们共进晚餐时,安德烈亚斯提到了他在业余时间从事的雄心勃勃的个人资助项目。他想要创建分辨率为 0.2 微米的冷冻小鼠的完整 3D 图,该分辨率足以绘制小鼠整个神经系统(包括大脑)的大部分神经连接。迈克·布朗没有出席颁奖典礼。他还离开了这家华尔街公司,此后在 Sun Microsystems 工作,然后加入了 Yago Systems,这是一家致力于网络连接智能交换机的成功初创公司。

Also present at the AAAI conference were several of our old friends from Deep Thought days, accepting the Allen Newell Award for Research Excellence for our joint work with Deep Thought. Thomas Anantharaman had left the Wall Street firm, and was working as an assistant professor at New York University, doing research related to the human genome project. Andreas Nowatzyk had left Sun Microsystems and joined Digital Equipment Corp. When we had dinner together after the ceremony, Andreas mentioned the ambitious, personally-funded project he was working on in his spare time. He wanted to create a complete 3-D map of a frozen mouse down to 0.2-micron resolution, a resolution good enough to map most of the neural connections of the mouse’s entire nervous system, including the brain. Mike Browne did not show up for the award ceremony. He had also left the Wall Street firm, had since worked at Sun Microsystems, and then joined Yago Systems, a successful startup working on smart switches for network connections.

移动的棋子

Moving Pieces

复赛后主角们都发生了什么?

What happened with the main characters after the rematch?

在早期的新闻文章中,据报道,胜利者“深蓝”已被拆卸并运往 IBM 客户,但没有国际象棋芯片。然而,它的死亡被夸大了。由于其历史意义,截至撰写本文时,深蓝仍然活跃在 IBM 研究中心。早期的内部报告称,只有深蓝的外框留在了IBM。事实证明,只有 RS/6000 节点回到了工厂测试车间,并被放回到原来的框架中。深蓝不再拥有复赛时使用的所有棋牌。两张卡片被送到位于 Armonk 的 IBM 总部,在那里设立了 Deep Blue Jr 来向参观者进行演示。大多数卡片仍然在 IBM 研究中心。1997 年的“深蓝”现在拥有原来棋牌的一半。一些卡片被放置在 1996 年比赛中使用的旧 RS/6000 SP 中。其中一些被放置在独立的工作站中。其余的都在货架上。

In early news articles, Deep Blue, the victor, was reported to have been disassembled and shipped to IBM customers, without the chess chips. Its death, however, was exaggerated. Due to its historical significance, as of this writing, Deep Blue is still alive and kicking in IBM Research. Early internal reports stated that only the outer frames of Deep Blue stayed in IBM. It turns out that only the RS/6000 nodes went back to the factory test floor, and have since been put back into the original frames. Deep Blue no longer has all the chess cards that it used in the rematch. Two cards went to IBM headquarters in Armonk, where a Deep Blue Jr was set up to give demos to visitors. Most of the cards are still in IBM Research. The 1997 Deep Blue now has half of its original chess cards. Some of the cards are placed in the old RS/6000 SP used in the 1996 match. A few are placed in standalone workstations. The rest are on shelves.

Deep Blue RS/6000 SP 主要用于与国际象棋无关的研究项目。它仍然可以下国际象棋——它作为一个网站可以同时与多达一千名国际象棋棋手对弈。然而,基于网络的程序甚至比“Pico”深蓝Jr(深蓝Jr的单芯片和减速版本)要弱得多。基于网络的程序使用单个深蓝国际象棋芯片,每步一秒,包括设置开销。每次移动的实际计算时间可能不到半秒。非常短的计算时间意味着基于网络的程序在战术上相对较弱,比典型的国际象棋程序在缩短计算时间时更弱。大多数对 Deep Blue Jr 的战术实力有部分影响的搜索扩展实际上都是不活跃的,考虑到计算时间短。该程序还使用简化的评估函数,不跟踪移动重复,也不“思考”对手的时间。它可能不会比 USCF 等级的 2000 级左右强。通常情况下,这很好,因为基于 Web 的程序的对手通常是 IBM 客户,并且让高兴的 IBM 客户在该程序中赢得几场比赛是完全可以接受的。1999年,国际计算机国际象棋协会要求IBM向计算机国际象棋比赛的观众提供基于网络的程序。一位商业国际象棋程序员利用了这个可怜的、毫无防御能力的基于网络的程序,通过与他自己的程序对弈,给他自己的程序带来了 50 比 1 的巨大时间优势。(为了公平地对待有问题的程序员,他可能然后,商业国际象棋程序员将他的程序赢得的比赛发布在他自己的网站上,作为他的程序“优越性”的证据,他非常高兴卖给你。

The Deep Blue RS/6000 SP is being used mainly for research projects unrelated to chess. It can still play chess—it serves as a web site that can play against up to one-thousand chess players simultaneously. The web-based program is, however, much weaker than even “Pico” Deep Blue Jr, the single chip and reduced speed version of Deep Blue Jr. The web-based program uses a single Deep Blue chess chip, at one second per move, including setup overhead. The actual computation time is probably less than half a second per move. The very short computation time means that the web-based program is comparatively weak tactically, more so than typical chess programs would be with shortened computation time. Most of the search extensions that are partially responsible for the tactical strength of Deep Blue Jr are effectively inactive, given the short computation time. This program also uses a simplified evaluation function, does not keep track of move repetition and does not “think” on the opponent’s time. It is probably no stronger than about 2000 rating on the USCF scale. This is fine, normally, as the opponents for the web-based program are usually IBM customers, and it is perfectly acceptable to have happy IBM customers winning a few games against the program. In 1999, the International Computer Chess Association asked IBM to provide the web-based program to the spectators in a computer chess event. A commercial chess programmer took advantage of the poor defenseless web-based program by playing it against his own program, giving his own program a hefty time advantage of the order of 50 to 1. (To be fair to the programmer in question, he probably did not know what the actual time advantage was when he played the games.) The commercial chess programmer then posted the games that his program won on his own web site as evidence of the “superiority” of his program, which he is perfectly happy to sell to you.

截至撰写本文时,自深蓝比赛以来,加里·卡斯帕罗夫还没有与任何计算机公开进行过严肃的比赛。他与人类玩家的比赛表现有戏剧性的起伏。

As of this writing, Garry Kasparov has not played serious games publicly against any computer since the Deep Blue matches. His performance against human players has its dramatic ups and downs.

在复赛后的一场锦标赛中,加里惩罚了他最早对阵的人类国际象棋棋手,并相当轻松地赢得了比赛。复赛后的一段时间内,他的棋绩有些参差不齐。他对深蓝的比赛的失败可能为他的人类对手提供了最初的心理动力。然后他连续不断地表现出色,国际象棋等级达到了 2800 分左右,远远高于任何其他人类国际象棋棋手。但在深蓝比赛中出现的盔甲裂缝最终又困扰着他。

In a tournament right after the rematch, Garry punished the first human chess players he played and won the tournament fairly easily. For a short while after the rematch, his chess results were somewhat spotty; his match loss against Deep Blue probably provided an initial psychological boost to his human opponents. Then he had a phenomenal run of great chess performances, with his chess rating going up to the middle 2800s, well above any other human chess player. But the cracks in his armor that showed up in the Deep Blue matches eventually came back to haunt him.

到2000年,加里已经五年多没有卫冕世界冠军了。正常的世锦赛周期是三年,最近的趋势是周期更短。近年来,国际象棋联合会国际棋联每年举办一次世界锦标赛。虽然没有人怀疑加里是最强的(人类)国际象棋棋手,但有人认为世界冠军头衔空缺,直到他与一位严肃的挑战者交手。自 1995 年与 Vishwanathan Anand 的冠军赛以来,他已经遇到了三位挑战者。第一个是阿列克谢·希罗夫 (Alexei Shirov),他赢得了他有权在利纳雷斯锦标赛上获得第二名来挑战他,领先于加里,但落后于阿南德,阿南德与国际棋联的书面合同阻止他在 1998 年参加任何非国际棋联世界冠军头衔。加里和阿列克谢之间的比赛从未进行过。1999年,加里在没有进行正常资格赛的情况下指定阿南德为新的挑战者,绕过了阿列克谢。阿南德当时世界排名第二,因此新的选择得到了一些公众的支持,尽管许多人不喜欢阿列克谢受到的对待方式,也不喜欢挑战者被任意选择的方式。原定与阿南德的新比赛也没有举行。2000年初,加里任命弗拉基米尔·克拉姆尼克为新的挑战者,取代阿南德。阿南德与比赛组织者就合同条件发生争执,认为他无法接受,结果被替换。卡斯帕罗夫与克拉姆尼克的比赛于2000年10月举行,弗拉基米尔在十五场比赛中以两胜十三平的成绩击败了加里,结束了加里十五年的统治,结果令人惊讶。阿南德并没有闲着,他自己也成为了国际棋联世界冠军。截至撰写本文时,弗拉基米尔 (Vladimir) 是 Braingames(2000 年比赛组织者)/爱因斯坦(2002 年从 Braingames 获得比赛权)世界冠军,18 岁的鲁斯兰·波诺马里夫 (Ruslan Ponomariov) 令人惊讶地成为 2002 年 FIDE 世界冠军,而加里 (Garry) 则是虽然没有获得世界冠军,但仍然是排名第一的选手,而弗拉基米尔则接近成为世界排名第二的选手。

By the year 2000, Garry had not defended his World Championship for over five years. The normal World Championship cycle was three years, and the recent trend is for even shorter cycles. In recent years, FIDE, the international chess federation, has held its own World Championship tournament once per year. While no one doubted that Garry was the strongest (human) chess player, there were people who considered the World Championship title vacant until he played a serious challenger. He had had three challengers since his 1995 title match against Vishwanathan Anand. The first was Alexei Shirov who won the right to challenge him by finishing second at the Linares tournament, ahead of Garry but behind Anand, whose written contract with FIDE prevented him from playing in 1998 for any non-FIDE world title. The match between Garry and Alexei never took place. In 1999, Garry designated Anand as the new challenger without a normal qualifying event, bypassing Alexei. Anand was ranked number two in the world, so there was some public support for the new choice, even though many people did not like the way that Alexei was treated nor the arbitrary way that the challenger was chosen. The scheduled new match with Anand also never took place. Early in 2000, Garry designated Vladimir Kramnik as the new challenger, replacing Anand. Anand had a dispute with the match organizer over contract conditions that he considered unacceptable, and was replaced as a result. The Kasparov vs. Kramnik match took place in October 2000, with the surprising result of Vladimir ending Garry’s fifteen-year reign by beating him in fifteen games with the score of two wins and thirteen draws. Anand did not sit idle and became the FIDE World Champion himself. As of writing, Vladimir is the Braingames (the match organizer back in 2000)/Einstein (the acquirer of match rights from Braingames in 2002) World Champion, 18-year-old Ruslan Ponomariov is the surprise 2002 FIDE World Champion, and Garry is without a world title but still the top ranked player, with Vladimir closing in as the second ranked player in the world.

自复赛以来,参与《深蓝》的大师们普遍都度过了美好的时光。为深蓝准备的开场白都没有被真正使用过,而宗师们也因为与深蓝的合作而拥有了非常强大的秘密开场武器。乔尔·本杰明和尼克·德弗米安分别成为 1997 年和 1998 年美国国际象棋冠军。当然,他们本身就是强大的宗师,无论是否与深蓝合作,都有很好的机会赢得冠军头衔。米格尔·伊列斯卡斯 (Miguel Illescas) 在弗拉基米尔 (Vladimir) 成功击败加里 (Garry) 的比赛中成为第二名,因此成为在加里 (Garry)两次比赛失利中唯一与对方球队有联系的人。

The Grandmasters who worked on Deep Blue have generally had good years since the rematch. None of the opening lines prepared for Deep Blue were really ever used, and the Grandmasters had very powerful secret opening weapons as a result of the cooperation with Deep Blue. Joel Benjamin and Nick DeFirmian became the 1997 and 1998 US Chess Champions respectively. Of course, they are strong Grandmasters in their own right, and had good chances to win the title with or without their work with Deep Blue. Miguel Illescas has the distinction of being Vladimir’s match second in his successful title bid over Garry, and thus becoming the only person to be involved with the opposing teams in both of Garry’s match losses.

IBM 的深蓝项目已经结束。默里和乔继续从事金融建模和数据挖掘,并取得了巨大成功。Joe 于 2001 年初离开 IBM 并加入了一家半导体初创公司,而我则于 1999 年底离开了 IBM。在他在 IBM 的最后一份计算机国际象棋工作中,Joe 修改了 Deep Blue,以提供一个简单的深蓝网络服务器的计算引擎。默里和乔以及我自己都曾就深蓝做过演讲并撰写过论文。然而,据我所知,穆雷和乔都不再从事电脑国际象棋的研究了。

The Deep Blue project had wound down in IBM. Murray and Joe moved on to financial modeling and data mining and had great success. Joe left IBM and joined a semiconductor startup in early 2001, while I left IBM earlier in late 1999. In his last computer chess work at IBM, Joe modified Deep Blue to provide a simple computation engine for the Deep Blue web server. Murray and Joe, as well as myself, have given presentations and written papers on Deep Blue. However, as far as I know, neither Murray nor Joe is working on computer chess any more.

改变景观

Changing Landscape

在我看来,加里不再公开打电脑比赛是他自己的选择。

In my opinion, it was Garry’s own choice that he did not publicly play a computer match again.

据我了解,IBM对与加里建立良好的赛后关系寄予厚望。无论深蓝是否获胜,都有可能与他合作的商业广告计划。他在赛后新闻发布会上的表现让这些计划被搁置。但据我所知,为了表现出良好的体育精神,IBM仍然愿意再给他一场比赛,至少在最初是这样。

From what I know, IBM had high hopes for a good post-match relationship with Garry. There were plans for possible commercial spots with him, whether Deep Blue won or not. His performance at the post-match press conference put those plans on hold. But as far as I can tell, to show good sportsmanship, IBM was still willing to grant him another match, at least initially.

复赛后的一段时间内,我研究了更强大的国际象棋芯片版本。一场新比赛的可能性以及消除卡斯帕罗夫阵营对他公平公正地输掉比赛的任何怀疑的愿望驱使着我继续前进。然后,有一天,我恍然大悟。“获得生活。你自由了。” 在我一生致力于这个项目的十二年里,我忘记了外面还有一个世界。

For a short while after the rematch, I worked on a more powerful version of the chess chip. The possibility of a new match and the desire to remove any doubt from the Kasparov camp that he lost fair and square, drove me on. Then, one day, it dawned upon me. “Get a life. You are free.” In the twelve years of my life I had dedicated to the project I forgot there was a world out there.

我知道我想做点别的事情;复赛后球队有很多公开活动,我好几个月都没有太多时间思考自己真正想做的事情。在此期间的某个时候,很明显 IBM 不会再与 Garry 进行任何进一步的比赛。虽然没有正式决定,但无休无止的阴谋最终造成了损失。

I knew that I wanted to do something else; the team had many public engagements after the rematch, and I did not have much time to think about what I really wanted to do for many months. Some time during this period, it became clear that there would not be any further IBM matches with Garry. There was no official decision, but the never-ending machinations finally took their toll.

当我终于有一些自己的时间时,我做出了三个决定。

When I finally had some time to myself, I made three decisions.

首先,我决定写这本书。虽然“深蓝”现在是一个家喻户晓的名字,但很少有人知道“深蓝”背后的真实人物。至少穆雷和我在计算机国际象棋界是众所周知的。乔的情况最糟糕。就连电脑象棋的人也对他了解不多。作为该项目的发起人,我相信我有责任澄清事实。此外,我确实有故事要讲。

First, I decided to write this book. While Deep Blue is now a household name, few know about the real people behind Deep Blue. At least Murray and I are well known in computer chess circles. Joe has the worst of it. Even computer chess people don’t know much about him. As the person who instigated the project, I believe that it is my responsibility to set the record straight. Besides, I do have stories to tell.

其次,我决定获得 IBM 的法律许可,以寻求在 IBM 之外进一步开发国际象棋芯片的可能性。鉴于 IBM 不会与 Garry 进行任何匹配,新的匹配发生的唯一方法就是我离开 IBM,并建造一台新机器。如果加里真的想要一场新比赛,我会尽我所能给他一场。如果他并不是真的想要一场新比赛,那么我会诚实地尝试回答他的挑战。

Second, I decided to get legal clearance from IBM to pursue the possibility of further developing the chess chip outside of IBM. Given that there would not be any IBM match with Garry, the only way a new match could take place would be for me to leave IBM, and build a new machine. If Garry really wanted a new match, I would do what I could to give him one. If he did not really want a new match, well, I would make an honest attempt to answer his challenges.

第三,一旦我重新获得了国际象棋芯片的控制权,我就会离开IBM。我想自己追求一些古怪的兴趣。看到安德烈亚斯在业余时间所做的事情让我想起尝试做一些不同寻常的事情是多么有趣。也许这是我的性格缺陷,但每隔一段时间我就喜欢把头伸进我不知道的专业领域的研究领域。在卡内基梅隆大学,我必须补习两门核心资格课程:计算机系统(硬件设计)1和人工智能。根据我的课程成绩,我没有资格成为像深蓝这样的机器的系统架构师。但这也是乐趣的一部分。我不知道我的古怪想法会走向何方,但这将是一次冒险。不管怎样,这就是我离开 IBM 的真正原因。从IBM获得国际象棋芯片的授权还要求我在签订合同后离开IBM。(这是任何许可谈判之前的先决条件。)但无论许可交易是否成功,我都会离开 IBM。

Third, I would leave IBM once I regained control of the chess chip. I wanted to pursue some off-the-wall interests on my own. Seeing what Andreas was doing in his spare time reminded me how much fun it was trying to do things out of the ordinary. Maybe it is a character flaw of mine, but every once in a while I like to poke my head into research areas in which I have no known expertise. At Carnegie Mellon, I had to take remedials for two of the core qualifier courses, in Computer Systems (hardware design)1 and Artificial Intelligence. Based on my course results, I had no business to be the system architect of a machine like Deep Blue. But that was part of the fun. I did not know where my off-the-wall ideas would lead, but it would be an adventure. Anyway, this was the real reason behind my leaving IBM. Licensing the chess chip from IBM also required that I left IBM after signing the contract. (It was the precondition before any licensing negotiation.) But I would have left IBM, whether the licensing deal went through or not.

我于 1999 年 10 月离开 IBM。经过一年多的合同谈判,我终于获得了法律许可,可以对棋子做或多或少的任何我想做的事情。我已经攒了足够的钱2来建造一台基于 PC 的国际象棋机器,如果我绝对需要的话,它比深蓝更强大。当然,如果可能的话,我更愿意将这笔资金保留下来用于我的其他兴趣。

I left IBM in October 1999. After more than a year of contract negotiation, I finally had the legal clearance to do more or less whatever I wanted with the chess chips. I had saved up enough money2 to build a PC-based chess machine that was more powerful than Deep Blue if I absolutely had to. Of course, if possible, I preferred to keep the fund in reserve for my other interests.

深蓝国际象棋芯片采用0.6微米CMOS技术制造。那时已经可以以合理的成本获得0.35微米甚至0.25微米的芯片。通过采用更先进的技术,完善棋片的设计,原本可以打造出一颗棋片3,其棋力几乎可以与整个深蓝相媲美。我必须像个疯子一样工作,但也有可能完毕。假设忽略与制造芯片相关的固定工程成本,有了一些这样的芯片,一台价格不到加里一天出场费 30,000 美元的十分之一的个人电脑就足以在一场比赛中击败他。

The Deep Blue chess chips were manufactured in 0.6 micron CMOS technology. It was possible to get 0.35-micron chips or even 0.25 micron chips at reasonable cost by then. By going to more advanced technology, and by refining the design of the chess chip, it would have been possible to create a single chess chip3 that would be almost comparable to the entire Deep Blue in chess playing strength. I would have to work like a mad man, but it could be done. With a few such chips, a PC costing less than one tenth of Garry’s one day appearance fee of $30,000, should be more than adequate to beat him in a match, assuming that the fixed engineering cost associated with fabricating the chip is ignored.

加里仍然公开挑战新比赛,甚至表示愿意将这场比赛打造成世锦赛。就我个人而言,我不认为应该允许机器参加世界锦标赛,但如果加里愿意,我不会与他争论。毕竟,他的意愿会让潜在的比赛赞助商更感兴趣。

Garry was still publicly making challenges for a new match and even went as far as suggesting that he would be willing to make the match a World Championship match. Personally, I didn’t believe that machines should be allowed to play for the World Championship, but if Garry was willing, I was not going to argue with him. After all, his willingness would have made potential match sponsors far more interested.

我的第一个行动是评估一些潜在比赛赞助商的兴趣。我得到的回应是,会有一些兴趣,特别是如果加里确实愿意将其设为世界冠军赛。这场比赛将成为历史性的第一场比赛。然而,尽管加里公开提出质疑,潜在的赞助商还是对他是否真的同意参加比赛表示严重怀疑。任何了解情况的人都清楚,IBM 将不再同意进行新的匹配。他一定非常清楚,在我离开 IBM 之前,他不必担心自己的挑战得到解答。另一个问题是,他总是可以说尚未制造的机器不是深蓝,尽管这台新机器代表着他必须再次与深蓝或类似的机器对战的唯一机会。我找不到任何论据来反驳发起人的推理。内心深处,我也对自己产生了一些怀疑。在确定他愿意参加比赛之前,赞助商不会采取任何行动。这种情况持续了大约两个月。此时,我向自己保证,如果他看起来不感兴趣,我就会开始追求我的其他兴趣。毕竟,我提出尝试安排比赛的唯一原因是为了回答他的挑战。如果他不认真对待自己的挑战,我为什么要费心呢?

My first action was to gauge the interests of a few potential match sponsors. The responses I got were that there would be some interests especially if Garry was indeed willing to make it a World Championship match. The match would then be an historic first. However, despite Garry’s public challenges, the potential sponsors expressed serious doubts that he would really agree to a match. It was clear to anybody with some knowledge of the situation that IBM would no longer agree to a new match. He must have known perfectly well that he did not have to worry about his challenges being answered, that is, until I left IBM. The other problem was that he could always say that the machine yet to be built was not Deep Blue, even though the new machine would have represented the only chance that he would ever have to play again against Deep Blue, or something close to it. I could not find any argument to refute the sponsors’ reasoning. Deep down, I had some doubts of my own. The sponsors would not do anything until they were sure that he was willing to play. This went on for about two months. At this point, I promised myself that if he did not seem interested, I would begin to pursue my other interests. After all, the only reason I was going through the motion of trying to set up the match was to answer his challenges. If he would not take his own challenges seriously, why should I bother?

在接下来的一个月左右的时间里,我与加里的经理欧文·威廉姆斯进行了几次接触。欧文立即排除了参加世界冠军赛的任何可能性。这对我来说很好。我对为冠军而战的想法感到不舒服。他对这场比赛不置可否。我什至无法让他说,原则上,如果能找到赞助商,加里愿意参加比赛。经过一个月的奔走,我得到了最终的确认加里不感兴趣。然后我向一些广泛阅读的国际象棋网站发出了一封公开信,并向欧文发送了一份副本。在信中,我说了再见,并向国际象棋界,包括加里本人表示感谢。我添加的信息表明加里对新比赛的挑战将不会得到答复,因为该挑战不再存在。我没有说为什么加里不再对这场比赛感兴趣,因为欧文没有给我任何关于他们是什么的暗示。

Over the next month or so I made several contacts with Garry’s manager, Owen Williams. Owen immediately ruled out any possibility of a World Championship match. This was fine by me. I was not comfortable with the idea of playing for the title. He was non-committal about the match; I could not even get him to say that, in principle, Garry was willing to play if a sponsor could be found. After a month of running around, I got final confirmation that Garry was not interested. I then sent out an open letter to a few widely read chess web sites and a copy to Owen. In the letter, I said my goodbyes, and thanks to the chess world, including Garry himself. I included information to say that Garry’s challenge for a new match would be unanswered, since the challenge no longer existed. I did not say anything about why Garry was no longer interested in the match as Owen had not given me any hint as to what they were.

在我的公开信几天后,欧文发表了他的回应,列出了加里对参加比赛不感兴趣的几个原因。其中原因就是我没有信誉,没有钱,我不认真。无论如何,进一步介入对我来说是双赢的局面。我发表了一个简短的回复,指出我不辞辛劳地获得了国际象棋筹码的权利,而且如果我绝对需要的话,我有足够的个人资金来建造一台新机器。我没有费心提及,我在 1985 年开始时没有计算机国际象棋资格,我们在没有任何预算的情况下建造了我们的第一台机器(ChipTest),但这并没有阻止深蓝在 1997 年击败加里。这完全是意志问题。也许整个过程中存在真正的误会,加里真诚地对比赛感兴趣。不管怎样,这个问题已经不再是我关心的了。我已经履行了尝试应对挑战的义务。

A few days after my open letter, Owen published his response that listed several reasons why Garry was not interested in playing a match. Among the reasons was that I had no credibility, no money, and I was not serious. Anyway, it was a no-win situation for me to get involved any further. I posted a short reply, pointing out that I went to the trouble to secure the right to the chess chip and I had enough personal funds to build a new machine if I absolutely had to. I did not bother to mention that I began with no computer chess credentials in 1985 and we built our first machine (ChipTest) with no budget whatsoever, which did not stop Deep Blue from beating Garry in 1997. It was all a matter of will. Maybe there were genuine misunderstandings in the whole process and Garry was sincerely interested in a match. Anyway, the issue was no longer of concern to me. I had fulfilled my obligation to try to answer the challenge.

打一场新比赛对加里来说没有什么好处,很可能会被压垮。计算机的进步比他快得多。1997年,一台价值数百万美元的超级计算机才击败了他。1999 年底,大学宿舍里的一台廉价 PC 上装有一些国际象棋芯片(用 0.35 微米 CMOS 制造)应该能够完成这项工作。在 2001 年超级碗电视转播中,百事可乐与加里一起出现。在广告中,他击败了一台看起来像哈尔(小说/电影“2001”中的疯狂计算机)的计算机,然后输给了百事可乐自动售货机。2001 年的现实是,像 Compaq iPaq 这样配备 0.13 微米 CMOS 国际象棋芯片的掌上电脑将有很大机会击败他。如果您想变得更奇特,您甚至可以将整个机器装入可乐罐中。这将是一个非常昂贵的罐头,但它是可以做到的。

There is nothing for Garry to gain by playing a new match, and it is very likely that he would be crushed. Computers can get better much faster than he can. In 1997, it took a multi-million-dollar supercomputer to beat him. In late 1999, a cheap PC in a college dorm with a few chess chips, fabricated in say 0.35 micron CMOS, should be able to do the job. During the 2001 Superbowl television broadcast, Pepsi ran a spot with Garry. In the commercial, he was seen beating a computer that looked like Hal (the insane computer in the novel/movie “2001”) and then losing to a Pepsi vending machine. The reality in 2001 is that a palmtop computer like a Compaq iPaq, equipped with a 0.13 micron CMOS chess chip would have a decent chance to beat him. If you want to get fancy, you can even fit the entire machine into a cola can. It would be one very expensive can, but it can be done.

由于加里目前没有头衔,谁可能是计算机的有趣对手?我的直觉是,弗拉基米尔可能是计算机更强大的对手。加里的比赛优势很大程度上在于他的战术打法,而这恰好是国际象棋计算机的强项。弗拉基米尔在比赛中压制加里战术的方式意味着他可能可以对计算机做同样的事情,这将使他成为一个非常危险的对手。

With Garry presently titleless, who might be an interesting opponent for computers? My gut feeling is that Vladimir could be a tougher opponent for computers. Garry’s playing strength resides to a great extent in his tactical play, which happens to be the forte of chess computers. The way Vladimir neutralized Garry’s tactical play in their match means that he probably can do the same thing to a computer, which will make him a very dangerous opponent indeed.

1997年在台湾台北举行的公开展览。

At a 1997 public exhibition in Taipei, Taiwan.

弗拉基米尔 (Vladimir) 和 Deep Fritz 节目之间的十场比赛原定于 2001 年 10 月进行,但被推迟。在撰写本文时,比赛已重新安排在 2002 年 10 月举行。如果您想知道,不,深弗里茨与深蓝没有任何关系。自从“深蓝”战胜了加里之后,一些商业程序就为其多处理器版本采用了“深蓝”这个名称。Deep Fritz 据说每秒能够搜索超过一百万个位置。基础程序 Fritz 的评估函数可能比 Deep Thought 的差。这意味着警惕的人类对手将能够利用弗里茨的位置弱点。弗拉基米尔应该轻松赢得比赛,除非他不小心。

A ten-game match between Vladimir and the program Deep Fritz was due to be played in October 2001 but was postponed. At the time of this writing, the match was rescheduled for October 2002. In case you are wondering, no, Deep Fritz has nothing whatsoever to do with Deep Blue. Ever since Deep Blue’s win over Garry, several commercial programs have adopted the name Deep “whatever” for their multiple-processor versions. Deep Fritz is supposedly able to search over a million positions/sec. The base program Fritz, has an evaluation function that is probably worse than Deep Thought’s. This means that an alert human opponent will be able to exploit Fritz’s positional weaknesses. Vladimir should win the match easily, unless he gets careless.


电脑游戏下一个感兴趣的研究领域是什么?下棋后,没有明确的选择。在具有完美信息的游戏(不依赖于抛硬币、掷骰子或洗牌等偶然事件的游戏)中,“围棋”通常被认为是最困难的游戏。这个问题可能非常困难,在未来 20 年内都无法“解决”。在其他游戏中,有两种特别令人感兴趣:中国象棋和将棋(日本象棋)。从技术上讲,中国象棋并不是很有趣,因为它的特点与国际象棋非常相似。中国棋手的数量可能与国际象棋棋手的数量相当。如果计算机首先在中国开发出来,那么这本书很可能是关于中国象棋计算机的。对我来说,中国象棋有一定的吸引力——这是我小时候学的第一个棋盘游戏。但我怀疑我是否会花时间在这上面。只是技术上还不够有趣。将棋是一个有趣的例子。它的复杂性介于围棋和国际象棋之间,主要来自于允许吃掉的棋子落回棋盘上的规则。我对如何解决将棋的计算复杂性有一些想法。有一天,我可能会建造一台将棋机,只是为了看看它有多难。这不会像《深蓝》项目那样需要全职工作——我想做一些实际的改变,而全职工作在另一款电脑游戏上是不可能的。我曾经热衷于下围棋,但我不相信我会继续下棋。目前这款游戏对于电脑来说太难了。

What will be the next research area of interest in computer games? After chess, there is no clear choice. Among games with perfect information (games that do not depend on chance events like coin flips, die rolls or card shuffling), “Go” is generally recognized as the most difficult game. It is probably sufficiently difficult that it will not be “solved” within the next 20 years. Of the other games, two are of particular interest: Chinese chess and shogi (Japanese chess). Technically, Chinese chess is not very interesting, since its characteristics are very similar to chess. The number of Chinese chess players is possibly comparable to the number of chess players. If computers had been developed in China first, then this book might very well have been about a Chinese chess computer. To me, Chinese chess holds some allure—it was the first board game I learned as a kid. But I doubt that I would spend time on it. It is just not sufficiently interesting technically. Shogi is an interesting case. It has a complexity between Go and chess, mainly from the rule that allows captured pieces to be dropped back onto the board. I have some ideas about how to tackle the computation complexity of shogi. Some day, I might build a shogi machine just to see how hard it is. It would not be a full-time effort like the Deep Blue project was—I want to do something practical for a change, and a full-time effort on another computer game is just not on the cards. I used to play Go passionately, but I don’t believe I will ever work on it. The game is too hard for a computer at the moment.

就我而言,电脑象棋已经不再是我生活的一部分了。我会同意建造一台新机器来对抗弗拉基米尔·克拉姆尼克或维什瓦纳坦·阿南德吗?我严重怀疑这一点,除非有人向我提出一个我无法拒绝的提议,因为我对新项目太感兴趣了。在我生命的这个阶段,对我来说,做一些新的、令人兴奋的事情比回到过去的征服更重要,尽管它可能很令人兴奋。我怀念竞争的快感吗?有时。但现在是我继续余生的时候了。这个世界充满了有趣的事情要做。

As far as I am concerned, computer chess is no longer a part of my life. Would I agree to build a new machine to play against Vladimir Kramnik or Vishwanathan Anand? I seriously doubt it, unless someone makes me an offer that I cannot refuse, as I am having too much fun with new projects. At this point in my life, it is far more important for me to do new and exciting things, rather than going back to an old conquest, exciting though it may be. Do I miss the thrill of competition? Sometimes. But it is time for me to get on with the rest of my life. The world is full of interesting things to do.

    好吧,我并没有那么糟糕。我直接参加了计算机系统课程的考试,没有参加课程本身或准备考试。

    Well, I was not that bad. I took the exam for the Computer Systems course directly without taking the course itself or studying for the exam.

    如果我没有足够的资源来构建新的国际象棋芯片,IBM 就不会允许我获得国际象棋芯片的许可。

    IBM would not have allowed me to license the chess chip if I did not have enough resources to build a new chess chip.

    芯片的运行速度应为 30-1 亿个棋局/秒,具体取决于所使用的技术以及每个芯片的国际象棋处理器数量。

    The chip should run at 30–100 million positions/sec, depending on the technology used and how many chess processors per chip.

 

 

2004 年平装本后记

AFTERWORD TO THE 2004 PAPERBACK EDITION

Braingames/Einstein 世界冠军弗拉基米尔·克拉姆尼克 (Vladimir Kramnik) 和无名第一选手加里·卡斯帕罗夫 (Garry Kasparov) 在 2002 年与计算机的比赛中都没有获胜。两场比赛都遵循相同的模式。人类一开始轻松赢得比赛,并以不止一场比赛领先。然后灾难降临了。他们都因不寻常的低级错误而倒下。在计算机将比赛扳平后,人类提出了早期的平局,计算机操作员立即接受了这些平局,两场比赛最终都以平局告终。

Neither Vladimir Kramnik, the Braingames/Einstein World Champion, nor Garry Kasparov, the titleless number-one player, won their 2002 matches against the computers. Both matches followed the same pattern. The humans won games easily at first and led by more than one game. Then disaster struck. They both fell after uncharacteristic rudimentary blunders. After the computers leveled the matches, the humans offered early draws that were promptly accepted by the computer operators, and both matches ended up drawn.

这两台计算机真的达到了弗拉基米尔或加里的水平吗?比赛比分就是这么说的。我相信吗?是和不是。弗拉基米尔和加里正在玩计算机的水平,但计算机并没有达到弗拉基米尔和加里的水平。

Were the two computers really playing at Vladimir or Garry’s level? The match scores said so. Do I believe that? Yes and no. Vladimir and Garry were playing at the computers’ level, but the computers were not playing at the level that Vladimir and Garry were capable of.

屡次的失误无疑表明人类并没有发挥出巅峰状态,而快速的平局表明两人在2002年的比赛结束时已经失去了斗志。尽管如此,计算机确实发挥了相当高的水平。问题是他们是否达到了世界冠军的水平。这个问题比较难回答。

The repeated blunders certainly indicated that the humans were not playing at their peaks, and the quick draws suggested that both had lost their fighting spirit by the end of the 2002 matches. Nonetheless, the computers did play at a reasonably high level. The question is whether they played at the world championship level. This is harder to answer.

我的直觉告诉我不会,但我并不能百分百确定。

My intuition says no, but I am not 100 percent sure.

甚至在 1996 年深蓝比赛之前,我就相信国际象棋知识对于世界冠军级别的国际象棋机器至关重要。我们的比赛第二名乔尔·本杰明(Joel Benjamin)在我们的实验室中反复证明,他可以利用自己卓越的国际象棋知识在对抗深蓝时发挥巨大作用,尽管深蓝的计算能力比他高。1996 年比赛的前四场比赛结束后,深蓝团队私下里私下议论:“乔尔在与电脑对战时会比加里更好吗?” 加里在随后两场比赛中的胜利表明也许并非如此。1997 版深蓝的主要开发工作是将其国际象棋知识提高到以前从未见过的水平。国际象棋计算机。我绝对确信它的国际象棋知识水平仍然超出了像 Deep Junior 和 Deep Fritz 这样的纯软件程序所能达到的水平。我也有理由相信深蓝在战术上比这两个程序都更好——比赛中的一些组合甚至连深蓝小号都抓住了,但两个程序却错过了。然而,深蓝真的需要全力发挥才能达到世界冠军的水平吗?对此我不确定。

Even before the 1996 Deep Blue match, I believed chess knowledge was crucial for world championship–level chess machines. Joel Benjamin, our match second, repeatedly demonstrated in our lab that he could use his superior chess knowledge to great effects against Deep Blue, despite the fact that Deep Blue could out-calculate him. After the first four games in the 1996 match, the Deep Blue team whispered among ourselves, “Could Joel be the better chess player than Garry against computers?” Garry’s wins in the two subsequent games suggested perhaps not. The main development effort for the 1997 version of Deep Blue was to improve its chess knowledge to a level never seen before in chess computers. I am absolutely sure that its chess knowledge level is still beyond what is possible on pure software programs such as Deep Junior and Deep Fritz. I am also reasonably sure that Deep Blue is better tactically than either program—there were combinations from the matches that even Deep Blue Jr caught but which were missed by the two programs. Yet, is Deep Blue’s full capability really needed to play at the world championship level? Of this I am not sure.

目前尚不清楚深蓝的超强战术能力是否必要。弗拉基米尔和加里所犯的错误根本不深,人类的胜利并不是真正基于深奥的计算,而是基于高超的国际象棋知识。所以,看来国际象棋知识仍然很重要。然而比赛还是平局了。为什么排名前两位的人不能将他们的国际象棋知识运用到其他比赛中呢?有几个简单的答案。一是他们没有做好充分准备。这是有道理的。如果这是真的,那么一些准备充分但水平较低的玩家应该相对容易地战胜这两个程序。另一个是他们被不输的压力压垮了。这也是情有可原的。两名人类玩家公开表示,他们的电脑对手比深蓝更强,但任何知情者都清楚,如果输掉比赛,他们就会丢面子。对我来说,最令人惊讶的答案是,也许当一个程序在战术上变得足够熟练时,人类就变得越来越难以利用其位置弱点,尽管这些弱点可能很明显。2002 年的比赛似乎表明,顶级人类利用程序的位置弱点的频率远远低于预期。我不喜欢这个答案——如果这是真的,那么我们就可以更早地完成深蓝项目。不过,比赛结果当然不排除这种可能性。

It is not clear that Deep Blue’s superior tactical capability is necessary. The blunders committed by Vladimir and Garry were not deep at all, and the wins by the humans were not really based on deep calculation but on superior chess knowledge. So, it appears that chess knowledge still matters. Yet the matches were drawn. Why couldn’t the top two humans bring their chess knowledge to bear in the other games? There are several easy answers. One is that they were not properly prepared. This is plausible. If this is true, then some better-prepared but lesser player should win relatively easily against the two programs. Another is that they were overwhelmed by the pressure not to lose. This is also plausible. Publicly, both human players stated that their computer opponents were stronger than Deep Blue, but it was clear to anyone in the know that they would have lost face if they lost the matches. To me, the most surprising answer is that perhaps when a program gets sufficiently proficient tactically, it becomes increasingly difficult for humans to exploit its positional weaknesses, blatant though the weaknesses may be. The 2002 matches seemed to indicate that the top humans could exploit the programs’ positional weaknesses far less often than expected. I don’t like this answer—if it is true, then we could have finished the Deep Blue project a lot earlier. However, the match results certainly did not preclude it.

在撰写本文时,加里·卡斯帕罗夫宣布他将于 2003 年 11 月 11 日至 18 日在纽约与 Deep Fritz(弗拉基米尔的前对手)进行一场表演赛。他还同意在没有实体棋盘的情况下进行比赛。相反,他将戴着 3D 眼镜并在虚拟棋盘上下棋。对于那些在深蓝比赛中要求完美照明条件的人来说,他同意这种不寻常的比赛条件显然给自己带来了严重的障碍。我祝他好运,但在这样的不利条件下,看到他因一系列失误而输得很惨,我也不会感到太惊讶。

At the time of this writing, Garry Kasparov announced that he would play an exhibition match against Deep Fritz (Vladimir’s former opponent) in New York City from November 11 to November 18, 2003. He has also agreed to play the match without a physical chessboard. Instead, he will be wearing 3D glasses and play on a virtual chessboard. For someone who requested perfect lighting conditions in the Deep Blue matches, he is obviously giving himself a severe handicap in agreeing to this unusual match condition. I wish him luck, but with this handicap, I would not be too surprised to see him lose badly through a series of blunders.

 

 

附录

Appendixes


 

 

附录A

APPENDIX A

来自台湾的小伙子

A Lad from Taiwan


这是背景材料,主要是关于我在台湾的生活。它应该让你了解我的核心是一个什么样的人,以及在讲述我的故事时可能存在什么样的偏见。《盒子里的生物》描述了我第一次接触国际象棋。学习棋盘游戏详细介绍了我是如何学习中国象棋、围棋和国际象棋的。它还提供了 20 世纪 70 年代之前的计算机国际象棋历史的简短入门。《早起者》是一篇关于我早期教育的短篇故事,也是我需要坦白的一个事件。回避权威提供了一些我早期倾向于忽视权威人物的例子。这种趋势与卡内基梅隆大学发生的事情有关。卡内基梅隆大学案例回顾了我最终选择卡内基梅隆大学的原因。

This is background material, mostly about my life in Taiwan. It should give you a sense of what kind of person I am at the core, and what kind of bias may exist in telling my stories. The Creatures in the Box describes my first contacts with chess. Learning Board Games details how I learned Chinese chess, Go, and chess. It also provides a short primer on computer chess history up to the 1970s. An Early Riser is a short story about my early education, and an incident that I need to come clean about. Sidestepping Authorities provides some instances of my early tendency to ignore authority figures. This tendency has a bearing on what happened at Carnegie Mellon. The Case for Carnegie Mellon goes over the reasons why I ended up at Carnegie Mellon.

盒子里的生物

The Creatures in the Box

我1959年出生于台湾基隆。基隆是台湾主要海港之一,当时人口约20万。在台湾,基隆被称为“雨港”,因为通常每年下雨超过200天。这是一座繁华的城市,但无论如何想象都不是一座迷人的城市。它也不是一座阴郁的城市;我在基隆看到彩虹的次数可能比世界上其他任何地方都多。

I was born in Keelung, Taiwan in 1959. Keelung is one of the major seaports of Taiwan, and at the time it had a population of around 200,000. In Taiwan, Keelung is known as the “Rainy Port” because typically it rains for more than 200 days a year. It is a bustling city, but not a glamorous city by any stretch of the imagination. It is not a gloomy city either; I probably saw rainbows more often in Keelung than anywhere else in the world.

当我上幼儿园时,我们的隔壁邻居买了附近的第一台电视机。由于台湾还没有彩色电视,所以是黑白的。每天下午,附近的孩子们都会围在一起看动画片。我想知道这些生物是如何进入盒子的。盒子后面有几个小洞,但它们太小了,生物无法通过。几周后,新鲜感逐渐消失,我们又回到了正常的比赛。

When I was in kindergarten, our next-door neighbor bought the first television set in the neighborhood. It was a black and white set, as there was no color television broadcast in Taiwan yet. Every afternoon, the kids in the neighborhood would gather around to watch cartoons. I wondered how the creatures got into the box. There were a few small holes behind the box but they were too small for the creatures to get through. After a few weeks, the novelty wore off, and we got back to our normal games.

几年后,我爸爸给家里买了一台电视机。当时,只有一家电视台,本地制作的电视节目几乎不存在。该台只在下午和晚上播出,但为了填补时段,最终重播了带有中文字幕的美剧。

A few years later, my dad bought a television set for the family. At the time, there was only one television station, and the local production of television shows was close to nonexistent. The station broadcast only on afternoons and evenings, but to fill the time slot, it ended up rebroadcasting American TV series, with Chinese captions.

当时国际象棋在美国并不盛行。鲍比·费舍尔 (Bobby Fischer) 赢得世界冠军带来的热潮来得稍晚一些。但有时,一些外国电视剧也会出现以国际象棋为主题的剧集。我第一次看到下国际象棋是在“家事”系列的一集中。对我来说,这是一个用精美的雕刻作品玩的奇怪游戏。

Chess was not big in the United States at the time. The boom resulting from Bobby Fischer winning the world title came a little bit later. But, occasionally, some of the foreign TV series carried episodes with chess themes. The first time I saw chess being played was in an episode of the “Family Affair” series. To me, it was a strange game played with beautifully sculptured pieces.

早期,这些进口电视剧大多是家庭情景喜剧。随着新电视台开始播出,推出了更多种类的节目。间谍节目变得相当受欢迎。《碟中谍》、《复仇者联盟》、《狂野西部》、《小偷来了》、《大叔的男人》、《变聪明》等等。1972 年费舍尔(美国)对斯帕斯基(苏联)世界象棋冠军赛的政治象征意义引起了编剧的注意,国际象棋被纳入了一些故事情节中。

In the early days, most of these imported TV shows were family sitcoms. As new stations began broadcasting, a wider variety of shows were introduced. Spy shows became quite popular. “Mission Impossible”, “The Avengers”, “Wild, Wild West”, “It Takes a Thief”, “The Man from U.N.C.L.E”, “Get Smart”, and so on. The political symbolism of the 1972 Fischer (USA) versus Spassky (USSR) World Chess Championship Match caught the attention of scriptwriters, and chess was incorporated into some of the story lines.

反复出现的电视主题之一是这样的。坏人是一位国际象棋大师,通常有东欧血统,而且非常聪明。我们的英雄出于某种原因必须与坏人进行一场国际象棋比赛,也许是为了接近坏人,营救另一位叛逃的大师,或者在国际象棋比赛中阻止一些险恶的事情。显然,我们的英雄最好打败坏人,从而让他犯下粗心的错误。有一个大问题。尽管我们的英雄通常与鲍比·费舍尔来自同一个国家,但他并不是国际象棋大师。当然,美国也以其计算机技术而闻名。因此,我们的英雄配备了一个隐蔽的远程连接到国际象棋计算机。对于我们的英雄来说,对抗坏人的游戏通常会顺利开始。为了达到戏剧性的效果,通常会在游戏进行到一半时出现问题。最后,要么计算机重新上线并消灭了坏人,要么我们的英雄偶然采取了将坏人的行动。镜头拉近坏人的脸,他完全不敢相信。现在受到惩罚的坏人冲出了大厅,在他引爆炸弹或做其他邪恶的事情之前,我们的英雄及时追上了他,冷酷地阻止了他。

One of the recurring TV themes went something like this. The bad guy is a Grandmaster chess player, usually of East European extraction, and very smart. Our hero has to play a chess match against the bad guy for some reason, perhaps to gain access to the bad guy, rescue another Grandmaster who is defecting, or stop something sinister at a chess competition. Obviously, it is preferable that our hero beat the bad guy, and thus upset him into making careless mistakes. There is one big problem. Even though our hero normally comes from the same country as Bobby Fischer, he is no chess Grandmaster. Of course, the USA is also known for its computer technology. So our hero is equipped with a well-hidden remote hookup to a chess computer. The game against the bad guy usually starts off fine for our hero. For dramatic effect, something typically goes wrong in the middle of the game. In the end, either the computer gets back on line and finishes off the bad guy, or our hero stumbles onto a move that checkmates the bad guy. The camera zooms into the face of the bad guy, who is in total disbelief. The now chastened bad guy then storms out of the hall, and before he can set off the bomb or do something else nefarious, our hero catches up with him, and stops him cold, just in the nick of time.

这个故事情节的寓意是什么?好吧,计算机是与坏人下棋的好工具,特别是如果您是一位英俊而温文尔雅的间谍1

What is the moral of this story line? Well, computers are great tools for playing the game of chess against bad guys, especially if you are a handsome and debonair spy1.

学习棋盘游戏

Learning Board Games

国际象棋在台湾并不流行。中国象棋和围棋(在西方,围棋的日语名称为围棋)都更受欢迎。这三者中,中国象棋最受欢迎。这也是我小时候学的第一个棋盘游戏。

Chess is not popular in Taiwan. Both Chinese chess and Wei-chi, which is better known as Go in the West from its Japanese name, are far more popular. Of the three, Chinese chess is the most popular. It was also the first board game I learned as a kid.

现在,如果这本书是关于某个著名国际象棋棋手的生活,我会告诉你我是如何通过观看我的父亲和哥哥下棋来学习的,有一天,我在他们刚刚完成的游戏中指出了他们的错误,让他们感到惊讶。那将是一个公然的谎言。事实并非如此。我在街上学到了中国象棋的规则。

Now, if this book was about the life of some famous chess player, I would be telling you about how I learned by watching my dad and older brother play, and one day surprised them by pointing out their mistakes in the game that they just finished. That would be a blatant lie. It was nothing like that. I learned the rules for Chinese chess on the street.

小时候基隆的汽车很少,邻居家的孩子在街上玩各种游戏。你可能会想,大半年都在下雨,我们怎么能在外面玩呢?那么台湾的商业楼宇,特别是基隆的商业楼宇一般都是两三层高,. 店面通常距离街道约十英尺。上层由街道旁边的柱子支撑,在商店前面形成一个悬挑。悬挑为顾客和路人提供了一个很好的庇护所。你可以从一栋建筑物的一端走到另一端而不会被雨淋到。下雨的时候,悬挑下面的空间就成了我们的游乐场。

There were very few cars in Keelung when I was little, and the kids in the neighborhood played all sorts of games on the streets. You might wonder how we could have played outside when it rained over half of the year. Well, the commercial buildings in Taiwan, especially in Keelung, were usually built two or three storeys high. The storefront was usually about ten feet from the street. The upper floor(s), supported by pillars next to the street, formed an overhang in front of the store. The overhang provided a nice shelter for the customers and the passers-by. You could walk from one end of a block of buildings to the other end without being rained on. The space under the overhang became our playground when it rained.

我第一次了解中国象棋的棋子是什么,是通过观看年龄较大的孩子玩更简单的游戏变体。每隔一段时间,他们中的一些人就会决定下“真正的”中国象棋。每当他们这样做时,我们这些年幼的孩子就会产生一种敬畏感。真实游戏中的棋子移动方式有所不同,我花了一些时间才大致了解真实游戏的玩法。我小学时的同学是我最早的对手,当我长大一点时,我爸爸也和我玩过一些游戏。大约之后,我变得有点和我爸爸竞争了。两周左右。我的一位叔叔来拜访我们并和我们一起玩了一些游戏。我叔叔比我们任何一个人都好得多,后来他慷慨地给了我一两点关于如何开始游戏和一般游戏原则的指导。这就是我接受的“正规”中国象棋教育的程度。

I first learned what the pieces for Chinese chess were by watching older kids play the simpler variants of the game. Every once in a while, some of them decided to play “real” Chinese chess. Whenever they did, there was a sense of awe among us younger kids. The pieces move differently in the real game, and it took some time before I had an inkling of how the real game was played. My classmates in primary school were my earliest opponents and my dad played some games with me when I got a little bit older. I became somewhat competitive with my dad after about two weeks or so. One of my uncles visited and played some games with us. My uncle was far better than either of us, and afterwards graciously gave a pointer or two on how to start the game, and general playing principles. That was about the extent of my “formal” Chinese chess education.

台湾学校的大部分中国象棋棋手的棋力显然相当薄弱。我初中时的打球实力就高于平均水平。一天下课后,一位朋友向我介绍了另一位中国棋手。我玩了几场比赛,完全被淘汰了。我很沮丧,之后我就很少打比赛了。

The playing strength of most of the Chinese chess players in Taiwanese schools apparently was quite weak. I was above average in playing strength when I was in junior high. One day after class, a friend introduced me to another Chinese chess player. I played a few games, and was completely outclassed. I was so discouraged that I played very few games thereafter.

在我学会下围棋之前,我确实学会了下西方象棋,或者更确切地说是一种混蛋形式的国际象棋。我的第一盘“国际象棋”是和我的哥哥凤龙下的。这是在他组装的自制国际象棋上进行的。当然,我哥哥赢了,部分原因是他随着游戏的进行告诉我规则。当我的兄弟升级他的棋子时,我在游戏接近尾声时学到了棋子升级2的规则。这个游戏大概是我上小学的时候玩的。初中的时候,我的一些同学偶尔会带着磁性象棋到学校,我们也确实下过一点“象棋”。我们谁都不太了解规则。尤其是我们的易位动作都是错误的。在中文中,易车易位被翻译为“交换城堡”,我们实际上是将王与车交换。

I did learn to play Western chess, or rather a bastardized form of chess, before I learned to play Go. My first game of “chess” was played against my elder brother, Feng-Lung. It was played on a homemade chess set that he put together. My brother won, of course, partly because of the fact he was telling me the rules as the game progressed. I learned the rule of pawn promotion2 near the end of the game, when my brother promoted his pawn. This game was probably played when I was still in primary school. In junior high, some of my classmates would bring their magnetic chess sets to school once in a while, and we did play “chess” a little bit. None of us knew the rules well. In particular, our castling moves were all wrong. In Chinese, the castling move is translated as “swapping the castle”, and we literally executed the move as a swap of the king with the rook.

我是在初中的时候开始对围棋产生兴趣的。作为学校课余娱乐活动的一部分,当地的业余三段选手时不时地过来给我们上课。除了礼仪、规则和一些基本模式之外,我们没有学到太多东西。当我考入台北市的正国中学时,我的一个姐姐景凤给我买了一套围棋作为礼物4。来自台湾的围棋职业九段林喜丰(日语:林海峰)在日本赢得了围棋大赛的主要冠军,台湾报纸广泛报道了他的冠军比赛。正是在这个时候,我发现了Go书籍。基隆并不是一个落后的城市,但它的文化程度却不如台北。基隆的书店很少,我家附近也没有。在台北,距离正国中学步行不远的地方,有一整条书店。Go 书籍改变了我后来处理技术问题的方式。每当我遇到新问题时,我的第一反应通常是挖掘我能找到的有关该问题的所有文献。在过去,这意味着去图书馆或书店。当然,今天我必须将互联网搜索添加到其中。有时,在查阅文献之前先思考一下问题会更好,但旧习惯很难改掉。

I started to get interested in Go when I was in junior high. As part of the school’s after-school recreational programs, a local amateur 3-dan3 player came in from time to time to teach us. We did not get much beyond the etiquette, rules and some basic patterns. When I got into Cheng-Kuo Senior High in Taipei, the capital, one of my elder sisters, Jing-Feng, bought me a Go set as a gift4. Professional 9-dan Lin Hi-Fong (Rin Kaihou in Japanese), the Go pro who hailed from Taiwan, was winning major Go titles in Japan, and Taiwanese newspapers were covering his title matches extensively. It was at this time that I discovered Go books. Keelung was not a backward city, but it was not as cultured as Taipei. There were few bookstores in Keelung, and none near our home. In Taipei, within walking distance from Cheng-Kuo Senior High, there was an entire street of bookstores. The Go books changed how I approached technical problems later on in life. Whenever I encounter a new problem, my first reaction is usually to dig up everything in the literature I can find about the problem. In the old days, this meant going to the library or bookstores. Of course, today I would have to add Internet searches to the mix. Sometimes, it is better to think about the problem first before looking it up in the literature, but old habits die hard.

学习 Go 在我的早期教育中非常重要。在远东,围棋不仅仅是一种游戏。它是文化的一个组成部分。在日本,商人有时会用围棋谚语来描述他们的商业策略。我对围棋的学习极大地影响了我的生活方式。我从 Go 中学到的最重要的教训之一就是选择自己的位置。在围棋中,对对手的每一次挑衅都做出反应通常是不明智的。有时,明智地忽视挑衅并转移到新领域会更有利可图。换句话说,做你自己,不要总是随波逐流。这种态度有时会适得其反,但总的来说,无论是在围棋游戏中还是在现实生活中,它都对我很有帮助。

Studying Go was quite important in my early education. In the Far East, Go is more than just a game. It is an integral part of the culture. In Japan, businessmen sometimes would use Go proverbs to describe their business strategies. My Go study greatly influenced how I approached life. One of the most important lessons I learned from Go was to pick my spot. In Go, it is frequently unwise to react to every single provocation from your opponent; sometimes, it is far more profitable to judiciously ignore the provocation and move to a new area. In other words, be yourself, and don’t always follow the crowd. This attitude sometimes backfires very badly, but it serves me well overall, in Go games and in real life.

学习围棋可能帮助我理解了国际象棋中的一些更深层次的概念。有时,当我向优秀的国际象棋棋手询问我从围棋文献中学到的特定形式概念时,答案常常是“是的,这是一个有趣的想法,我确实直观地使用了这个概念”。然而,这个概念在国际象棋文献中通常没有被命名。围棋是一个复杂的游戏,但它的规则非常简单。也许 Go 规则的简单性使得更容易明确地描述抽象概念。

Learning Go probably helped me understand some of the deeper concepts in chess. Sometimes when I asked good chess players about a specific formal concept that I learned from the Go literature, the answer frequently was “yes, that is an interesting idea, and I do use the concept intuitively”. The concept, however, is usually not named in the chess literature. Go is a complicated game, but it has very simple rules. Perhaps the simplicity of Go rules makes it easier to describe the abstract concepts explicitly.

我在大一的时候就学会了国际象棋的正确规则在国立台湾大学。有一天,我决定在学校图书馆的目录中查找有关国际象棋的书籍,并惊讶地发现列出了一本计算机国际象棋书籍。我在电气工程系,所以我对计算机有一些兴趣。找到一本列出的计算机国际象棋书籍是相当令人兴奋的,尤其是目录上列出的作者正是世界国际象棋冠军米哈伊尔·鲍特维尼克。当我找不到他的书时,我很生气。

I learned the proper rules for chess during my freshman year in National Taiwan University. One day, I decided to look through the catalogue of the school library for books on chess and was surprised to find a computer chess book listed. I was in the Electrical Engineering Department, so computers were of some interest to me. Finding a computer chess book listed was quite exciting, especially with the catalogue listing the author as none other than the World Chess Champion Mikhail Botvinnik. I was very annoyed when I could not locate his book.

事后看来,没有找到这本书可能是因祸得福。他所提倡的方法似乎毫无进展。大约二十年来,他写了许多关于他的程序在测试位置上表现如何的文章。但是,直到今天,还没有外部研究人员看到他的程序下过棋,更不用说与其他程序竞争了。一些研究人员甚至指责他在考试职位上伪造成绩。他有一些好的想法,至少根据我交谈过的国际象棋棋手的说法,但即使在更强大的西方计算机上,这些想法也很难实现。在铁幕背后,它们很可能无法实施。他于 20 世纪 90 年代初去世,当时正值个人电脑在前苏联广泛普及。

With hindsight, not finding the book might have been a blessing in disguise. The approach he advocated appeared to go nowhere. For twenty years or so, he wrote many articles on how well his program performed in test positions. But, even to this day, no outside researcher has seen his program playing a single move, let alone competing against other programs. Some researchers even accused him of falsifying his results in the test positions. He had some good ideas, at least according to the chess players to whom I have spoken, but the ideas were difficult to work with even on the more powerful Western computers. Behind the iron curtain, they were likely impossible to implement. He died in the early 1990s, right around the time when the personal computer became widely available in the former USSR.

不管怎样,随着我的兴趣的激发,我浏览了所有我能找到的关于计算机国际象棋的期刊文章。我学习国际象棋的规则,不是从书本上,而是从一篇计算机国际象棋文章中学到的,这篇文章恰好详细描述了规则。

Anyway, with my interest piqued, I went through all the journal articles that I could find on computer chess. I learned the rules of chess, not from a book, but from one of the computer chess articles that happened to give a detailed description of the rules.

我继续时不时地查看图书馆目录。在我大二的时候,一本新的电脑国际象棋书籍到达了学校图书馆。这本书是计算机国际象棋研讨会的论文集,名为《人与机器的国际象棋技巧》,现在被计算机国际象棋研究人员认为是经典。本书中最重要的一章是由 David Slate 和 Lawrence Atkin 编写的国际象棋程序Chess 4.5。(大多数现代国际象棋程序都以某种方式克隆了 Chess 4.5 程序。)当我读这本书时,我不知道它会产生多大的影响力。我最惊讶的是它的悲观基调,特别是考虑到媒体通常乐观的报道,但事实证明悲观主义要现实得多。

I continued to check the library catalogue from time to time. During my sophomore year, a new computer chess book arrived in the school library. The book, a collection of papers from a computer chess workshop, was Chess Skill in Man and Machine, which is now considered a classic by computer chess researchers. The most important chapter in the book is on the chess program Chess 4.5 by David Slate and Lawrence Atkin. (Most of the modern chess programs are in one way or another clones of the Chess 4.5 program.) At the time I read the book, I had no idea how influential it was to become. I was mostly surprised by its pessimistic tone, especially given the usually rosy reports in the press but the pessimism turned out to be far more realistic.

计算机国际象棋文献普遍将克劳德·香农 1949 年的演讲视为现代计算机国际象棋历史的开端。一些早期的研究人员相当乐观。1957年,诺贝尔经济学奖获得者赫伯特·西蒙(Herbert Simon)著名计算机科学家甚至声称“十年内数字计算机将成为世界国际象棋冠军,除非规则禁止它参加比赛”。十年后,最好的程序是由麻省理工学院的学生理查德·格林布拉特编写的。他的程序处于 C 级水平的低端,按照国际象棋评级标准,为 1400 到 1600 分。在相同等级上,世界冠军的评分通常在 2800 到 2900 左右。1400的选手比初学者要好,但会输给超过一半的周末锦标赛选手,更不用说世界冠军了。比对手强一级(高 200 评级点)的玩家应该在每四场比赛中赢得大约三场。1967 年最好的计算机国际象棋程序的成绩为 1400 分,即比世界冠军低 7 个等级。赫伯特·西蒙的预测大错特错。

The computer chess literature generally regards Claude Shannon’s 1949 lecture as the beginning of modern computer chess history. Some of the early researchers were quite optimistic. In 1957, Herbert Simon, a Nobel laureate in Economics, and renowned computer scientist, went as far as stating “within 10 years a digital computer will be the world’s chess champion, unless the rules bar it from competition”. Ten years later, the best program was authored by Richard Greenblatt, an MIT student. His program played at the low end of Class C level which, on the chess rating scale, would be 1400 to 1600 rating points. World Champions are typically at around 2800 to 2900 rating points on the same scale. A 1400 player is better than a beginner, but would lose to more than half of the weekend tournament players, let alone the World Champion. A player one class stronger (200 rating points higher) than their opponent should win about three games out of every four. The best computer chess program in 1967 was playing at 1400 points, or seven classes, below the World Champion. Herbert Simon’s prediction was way off the mark.

1975 年,当《人与机器的国际象棋技能》的大部分材料编写完毕时,Chess 4.5 是最好的程序。然而,其程序员始终坚称该程序仍处于 C 级水平,即比理查德·格林布拉特 (Richard Greenblatt) 的水平高出不超过 200 分。换句话说,最好的国际象棋程序经过近 10 年的开发,得分还不到 200 分。按照这样的进展速度,计算机还需要 60 年,或者说 2035 年左右,计算机才能挑战世界冠军。难怪这本书的总体基调是悲观的。几年后,当在超级计算机上运行的 Chess 4.9 被证明可以达到接近 2000 个评级点时,一些乐观情绪又回来了。但直到多年后我在卡内基梅隆大学攻读博士学位时,我才知道这种新的乐观情绪。

In 1975, when most of the material in Chess Skill in Man and Machine was written, Chess 4.5 was the best program. Its programmers, however, were consistently maintaining that the program was still at Class C level, or no more than 200 points higher than Richard Greenblatt’s. In other words, the best chess programs gained less than 200 points after almost 10 years of development. At this rate of progress, it would have taken 60 more years, or around the year 2035, before a computer could challenge the World Champion. No wonder the general tone of the book was pessimistic. Some optimism came back a few years later when Chess 4.9, running on a supercomputer, was shown to play at close to 2000 rating points. But I did not know about the renewed optimism until I was studying for my PhD degree at Carnegie Mellon University many years later.

在台湾很难找到象棋书籍。那些都是英文的,我必须去专门的书店才能找到。我找到的要么是游戏合集,要么是入门书籍。入门书籍太初级了,没有超出游戏规则,游戏宝典也没有解释我这个水平的东西。我第一次接触国际象棋位置概念是来自计算机国际象棋文献。直到我在美国时,我才读到第一本关于国际象棋走位概念的书。当我搬到美国时,我的国际象棋水平非常差。

Chess books were difficult to find in Taiwan. The ones available were in English and I had to go to specialty bookstores to find them. Those that I found were either game collections or introductory books. The introductory books were too elementary—they did not go much beyond the game rules, and the game collections did not explain things at my level. My first contacts with chess positional concepts were from the computer chess literature. It was only when I was in the United States that I read my first book on chess positional concepts. I was a very weak chess player when I moved to the United States.

在大学期间,我时不时地继续下围棋。我也下了几盘棋,主要是因为我对电脑象棋的兴趣。我几乎没下过中国象棋,尽管我确实记得当我尝试“像计算机一样思考”时,通过想象程序如何“推理”该位置,赢得了一场休闲游戏。然而,我在那场中国象棋比赛中的思维过程与几十年后深蓝实际上是如何下棋的没有任何关系。

During my college years, I continued to play Go from time to time. I also played a few games of chess, mostly because of my interest in computer chess. I hardly ever played Chinese chess, although I do remember winning one casual game when I tried to “think like a computer”, by imagining how a program would “reason” about the position. My thought process in that Chinese chess game, however, had no bearing whatsoever on how Deep Blue actually played chess decades later.

大学毕业后,我服了两年义务兵役,然后去了美国卡内基梅隆大学。我在美国的时候就不再下围棋了。当我退出时,我已经达到了美国标准的初段(一段)。退出围棋大约一年后,我在卡内基梅隆大学开始了自己的计算机国际象棋项目,但这两个决定完全无关。我从来都不是真正的国际象棋棋手,即使在我开始全职从事电脑国际象棋工作之后,也很少下棋。由于计算机国际象棋工作,我的国际象棋水平变得更好,但我从未有过正式的国际象棋评级。我很确定我的国际象棋评分充其量也不过 1800 分。我从深蓝项目中真正学到的是对大师们所实践的艺术之美有了更强烈的欣赏。

After college, I spent two years in mandatory military service, and then went to Carnegie Mellon University in the United States. I quit playing Go when I was in the US. I was about shodan (one dan) on the American scale when I quit. I started my own computer chess project at Carnegie Mellon about a year after quitting Go, but the two decisions were totally unrelated. I was never really a chess player, and played few games even after I began working full time on computer chess. I became better at chess because of the computer chess work, but I have never had a formal chess rating. I am pretty certain that I was, at best, no better than an 1800 on the chess rating scale. What I really learned from the Deep Blue project is a stronger appreciation of the beauty of the art practiced by the Grandmasters.

早起者

An Early Riser

下面的故事与国际象棋或电脑象棋无关,但它确实告诉了你一些关于我的事情。我把它放在这里的原因是为了澄清事实。我的父母、我的老师和我的初中同学都应该知道真相。

The following story has nothing to do with chess or computer chess, but it does tell you something about me. The reason why I include it here is to come clean. My parents, my teacher and my classmates in junior high deserve to know the truth.

台湾的教育制度可以追溯到中国唐朝。在此期间,一位皇后在她的皇帝丈夫去世后接管了中国的统治,成为第一位也是唯一一位直接统治国家的皇后。虽然传统主义者可能不会看好她的统治,但她的一项法令对中国人的心理以及整个远东的家庭传统产生了深远的影响。她建立了基于公开考试的选拔政府官员的制度。公开考试成为主要的社会移民途径,因为对谁可以参加没有限制。最穷农民的儿子有可能成为总理。对于远东的许多父母来说,孩子的教育成为家庭生活中最重要的事情,因为该地区的其他统治者效仿了这一制度。随着时间的推移,教育体系本身开始围绕各种考试展开。这种具有千年历史的教育制度,旨在为朝廷培养听话的文官。当然,创造力并不一定是服从的文官的理想品质。

The education system in Taiwan can be traced back to the Tang Dynasty in China. During this period, an Empress took over the reign of China after the death of her Emperor husband, becoming the first and only Empress to directly rule over the country. While traditionalists might not look upon her reign favorably, one of her edicts had profound effects on the Chinese psyche, and the family tradition of the entire Far East. She established a system for selecting government officials based on public exams. The public exams became the main social migration paths as there was no restriction on who could take part. Potentially, the son of the poorest peasant could become the Prime Minister. Education of children became the most important thing in family life for many parents in the Far East as other rulers in the region copied the system. In time, the education system itself began to revolve around examinations of various sorts. This thousand-year-old education system was designed to produce obedient civil officials for the imperial government. Creativity was, of course, not necessarily a desirable trait for obedient civil officials.

我上小学的时候,台湾的一个孩子要经过三门考试才能上大学。第一次入学考试是初中(小学六年后),第二次是高中,最后一次是高考。前两次是地区性的,但高考是全国性的。在我三年级的时候,台湾政府决定所有孩子都应该接受至少九年的教育,并且废除了初中入学考试。所以我确实度过了一段没有立即入学考试压力的童年。

At the time when I was enrolled in primary school, a Taiwanese youngster needed to go through three entrance exams before reaching college. The first entrance exam was for junior high (after six years of primary school), the second was for senior high, and the last was the college entrance exam. The first two were regional, but the college entrance exam was nationwide. During my third grade, the Taiwanese government decided that all children should have at least nine years of education, and the junior high entrance exam was abolished. So I did have a little bit of childhood without the immediate pressure of the entrance exams.

高考重记忆,导致台湾整个教育体系严重扭曲。台湾的许多学生在正规学校放学后都会花时间参加私人补习班,为入学考试而学习,这给父母带来了相当大的费用。说实话,我不知道他们在那些私人课程中学习了什么或如何学习,因为我自己从未参加过。我设法避免上私人课程,因为我在小学时就养成了一种独特的“学习”方法。我的“学习”方式只利于死记硬背,并不利于真正理解学科。有一些科目,比如数学和物理,仍然需要定期学习,但其他只需要记忆的科目我就差不多了。

The entrance exams placed a premium on memorization, and the whole education system in Taiwan was seriously skewed as a result. Many students in Taiwan would spend time in private tutoring classes after regular school, studying for the entrance exams at considerable expense to their parents. To tell the truth, I have no idea what or how they studied in those private classes, as I never attended one myself. I managed to avoid taking private classes because I developed a peculiar method of “studying” while I was in primary school. My way of “studying” was only good for memorization, and not for truly understanding the subject. A few subjects, such as math and physics, still required regular study, but I pretty much got by with the other subjects that only required memorization.

我弟弟上初中,我还在上小学。有时候,考试那天他会起得很早,因为他前一天的学习还没有完成。但当我想到这一点时,在考试当天进行学习是完全合理的,假设它可以完成。有几个很好的理由。

My brother was in junior high while I was still in primary school. Occasionally, he would get up very early on the day of school exam because he had not finished studying from the day before. But when I thought about it, it made perfect sense to do the studying on the very day of the exam, assuming that it could be done. There were several good reasons.

首先,我的记忆依然清晰,不会因一夜睡眠而衰退。其次,学习时间短会迫使我集中注意力,因此实际需要的学习时间会减少。第三,台湾的亚热带气候,清晨学习要舒服得多。当然,我真正的动机是,如果这个计划有效,我就不需要花很多时间学习了。我是一个懒惰的人,而且我的懒惰很早就开始了。

First, my memory would still be fresh, without the decay from a night of sleep. Second, the short time available for studying would force me to focus, and thus less studying time was actually needed. Third, given Taiwan’s subtropical weather, it was far more pleasant to study in the early morning. Of course, my true motivation was that if the scheme worked, I would not need to spend a lot of time studying. I am a lazy person, and my laziness started very early in life.

到了初中的时候,早起的临时抱佛脚已经成为我的一种习惯。学校考试那天,我会在凌晨 1 点起床,学习到天亮。服用后考试时,我崩溃了,忘记了早上强行灌入大脑的所有愚蠢的花絮。我的成绩相当不错,一直在班上名列前茅或接近前茅。我的老师开始注意到了,因为我是班上唯一一个没有上私人课程的尖子生。

By the time I was in junior high, early morning cramming had become a ritual for me. On the day of a school exam, I would get up at, say 1.00 am, and study until daybreak. After taking the exam, I crashed and forgot all the silly tidbits that I force-fed into my brain in the early morning. I was getting fairly good grades, constantly at or near the top of the class. My teacher started to take notice, as I was the only student at the top of the class not taking private classes.

一天早上,我的老师向全校发表了演讲。那天早上我值班打扫教室,幸好没有出席演讲。后来我很震惊地发现我就是那部分演讲的主题。显然,我的老师听说了我在凌晨学习的事实,并热情地讲述了其他学生应该如何效仿我的做法。他假设我每天都这样做,并以我为例,告诉其他学生应该如何像我想象的那样努力学习。

One morning, my teacher gave a speech to the entire school. That morning I was on duty to clean up the classroom, and thankfully was not present for the speech. I was shocked to find out later that I was the subject of part of the speech. Apparently, my teacher got wind of the fact that I was studying in the wee hours of the morning, and gave a glowing account of how the other students should follow my example. He was assuming that I did it every day, and was using me as an example on how the other students should study as hard as I supposedly did.

我的老师从来没有意识到我早起学习的真正原因。更让我尴尬的是,妈妈听了同学妈妈的话,得意地重复了老师说的话。我从未告诉我的母亲或老师到底发生了什么。好吧,现在他们知道了。

My teacher never did realize the real reasons for my early morning studying. I was even more embarrassed when my mother proudly repeated what my teacher said after hearing it from my classmates’ mothers. I never did tell my mother or my teacher what really transpired. Well, now they know.

回避权威

Sidestepping Authorities

众所周知,青少年是叛逆的。我的行为还算不错,但我确实有叛逆的倾向。当当局以令我满意的方式证明他们值得尊重时,我尊重他们,但尊重当局并不意味着我不会试图回避他们。这种对权威人物的态度对卡内基梅隆大学后来在国际象棋项目的早期发生的一些事件产生了一些影响

Teenagers are known to be rebellious. I was reasonably well behaved, but I did have a rebellious streak. I respected the authorities when they proved to my satisfaction that they deserved respect, but respecting the authorities did not mean that I would not try to sidestep them. This attitude towards authoritative figures had some bearing on some of the later events at Carnegie Mellon, during the early years of the chess project

我上高中时,我哥哥正在大学学习化学工程,我捡到了他不再使用的一些大学教科书。我高中的第一位数学老师是一个相当不负责任的人,他在课堂上一半以上的时间都在抱怨这抱怨那,而不是真正的教学。听了他几次后,我决定我在课堂上的时间最好花在自己的阅读上,所以我开始翻阅我哥哥的数学教科书。数学老师忙着抱怨他的生活,他从来没有意识到我在读自己的书,或者也许他只是不在乎。我的第二位数学老师是一位好老师,但由于我已经浏览了一些材料,所以我时不时地在课堂上继续自己的课外数学学习。一天老师发现我没有注意他,就对我下手了。当他发现我在读什么时,他退缩了,但警告我不要再这样做了。毫不奇怪,他开始挑剔我,经常要求我走到黑板前证明我知道这个主题。这种情况持续了一段时间,直到他提出了一个他认为我不可能解决的问题。

My brother was studying Chemical Engineering in college when I enrolled in senior high, and I picked up some of his college textbooks that he was no longer using. My first math teacher at the senior high was a fairly irresponsible person, who spent more than half of the time in class complaining about this and that instead of doing real teaching. After listening to him a few times, I decided that my time in class would be better spent doing my own reading, so I started going through my brother’s math textbooks. The math teacher was so busy complaining about his life that he never realized that I was reading my own books, or maybe he just did not care. My second math teacher was a good teacher, but since I had already gone through some of the material, I continued my own extracurricular math study in the class from time to time. One day the teacher noticed that I was not paying attention to him and came down on me. When he found out what I was reading, he backed off but gave me a warning not to do it again. Not surprisingly, he began to pick on me, frequently asking me to go up to the blackboard and demonstrate that I knew the subject. This went on for a while until he came up with a problem that he thought I could not possibly solve.

结果是求解由随机游走问题产生的二阶有限差分方程5 。当然,有限差分方程实际上并不是高中课程的一部分。事实上,它们也不在我的大学课程中。但我很幸运。它们出现在我哥哥的大学课程中,几个月前我碰巧在他的书中读到过它们。我走到黑板前,用了大约 30 秒的时间解出了这个方程,所用的符号是其他学生都不知道的。看到老师的脸是无价的。我知道他不可能指望我用我所用的方法解方程。我很好奇他将如何向全班解释事情,以及他的替代解决方案是什么。

It turned out to be solving a second order finite difference equation5 resulting from a random walk problem. Of course, finite difference equations are not really part of the high school curriculum. As a matter of fact, they were not in my college curriculum either. But I got lucky. They were in my brother’s college curriculum, and I happened to have read about them a few months earlier in his books. I went up to the blackboard and solved the equation in about 30 seconds, using notations none of the other students knew anything about. It was priceless to see the teacher’s face. I knew that he could not possibly have expected me to solve the equation using the method that I used. I was curious about how he was going to explain things to the class, and what his alternative solution was.

“嗯,这被称为有限差分方程,但我们不会那样解决它。” 所以他确实有另一种方法来解方程。这会很有趣。接下来的十几分钟,他用复杂的代数方程写满了半块黑板,最终得到了和我刚才一样的解。后来查了一下补充书籍,发现确实是里面推荐的解决方案。

“Well, that was called a finite difference equation, but we are not going to solve it that way.” So he did have an alternative way to solve the equation. This was going to be interesting. For the next ten minutes or so, he filled up half of the blackboard with complicated algebraic equations, and eventually obtained the same solution as I had earlier. I checked the supplementary books afterwards, and found out that was indeed the solution recommended in them.

事情发生后,老师对我的态度就放松了。我给我的一些高中同学留下了深刻的印象,他们给我起了个绰号“疯狂的小鸟6 ”。后来,当我来到美国时,那些读不出我名字的人都知道我是CB。

The teacher eased up on me after the incident. I made quite an impression on some of my high school classmates, and they gave me the nickname “Crazy Bird6.” Later on, when I came to the United States, people who had problems pronouncing my name knew me as CB.

我考入台大后,和一些教授发生过一些小事,但都没有什么严重的。一事件发生在物理课上,和我高中数学课上发生的事情有些相似,但没有那么戏剧性。后来,我们的物理学教授产生了一个奇怪的想法:我会成为一名优秀的物理学家。我成绩很好,所以我没有理由抱怨。我在语文课上做了一些更大胆的事情。我们应该根据我们在中国文学经典《传子》中读到的内容写一篇中文文章。《传子》恰好与中文的“原子”押韵。这篇文章的题目是《论读《传子》》。我把标题改成《传子碰撞实验》,以原子碰撞实验报告的形式写成一篇讽刺文章。在我的文章中,我广泛地取笑了我们的中国教授。他对我对他的描述不太满意,但也很开心地给了我一篇论文的最高分。正如我所怀疑的,他确实有很强的幽默感。

After I enrolled in the National Taiwan University, I had a few small incidents with some professors, but nothing serious. One incident took place in the physics class and was somewhat similar to what happened in my high school math class, but not as dramatic. Afterwards, our physics professor had the strange notion that I would make a good physicist. I got good grades, so I had no reason to complain. I did something more daring in my Chinese class. We were supposed to write a Chinese essay on what we read in Chuan-Tzu, a Chinese literary classic. Chuan-Tzu happened to rhyme with the Chinese translation for “atomic”. The title for the essay was something like “On Reading Chuan-Tzu”. I changed the title to “Experiments on Chuan-Tzu Collisions” and wrote the essay as a satire in the form of a report on atomic collisions experiments. In my essay, I made fun of our Chinese professor extensively. He was not too happy with my portrayal of him, but was sufficiently amused to give me my highest ever grade for an essay. As I suspected, he did have a strong sense of humor.

在台湾,男孩大学毕业后必须服兵役两年。在我看来,最终的结果是大学男生变得更笨了。我当然觉得自己更蠢了。我的反应速度较慢,一路上我肯定失去了优势。军事环境的目的是培养听话的士兵,而不是培养自由思想的灵魂。缺乏智力刺激导致智力萎缩。事实上,我们许多人都处于高压环境中,这也无济于事。我比大多数人都幸运,因为我的服务主要是在导弹指挥训练中心教授新士兵和下级军官,基础电力和英语科目。时间很长,令人沮丧的是,大多数士兵学生都是来自桶底。大多数学生要么不能,要么不愿意学习英语字母。最后,我只专注于少数有机会成功的人。无论如何,我不高兴,我想改变一切。

In Taiwan, boys were required to serve two years in the military after college. In my opinion, the net effect was that college boys became dumber; I certainly felt dumber. My reaction time was slower, and I definitely lost an edge along the way. The military environment was designed to produce obedient soldiers, not free thinking souls. The lack of intellectual stimulation led to mental atrophy. The fact that many of us were placed in high-pressure environments also did not help. I was luckier than most, in that my service was mainly teaching new soldiers and junior officers, the subjects of basic electricity and English, at the Missile Command Training Center. The hours were long, and it was frustrating that the majority of the soldier students were from the bottom of the barrel. Most of the students either could not or would not learn the English alphabet. In the end, I just concentrated on the few that had some chance of making it. Anyhow, I was not happy, and I wanted to change things.

导弹司令部有一些模拟器,用于训练负责发射导弹的军官。模拟器很贵,而且都坏了。在成为教官之前,我接受过指挥舱的维护培训,对指挥舱的内部结构有些熟悉。该模拟器基于修改后的命令模块,实际上是一款使用雷达屏幕作为显示器的美化视频游戏。该中心拥有模拟器和命令模块的原理图。它们基于古老的技术:基于继电器的逻辑、用于波形生成的模拟电位器、等等。典型的军事装备,已经过时十到二十年了。命令模块的视频流很容易访问,并且修改它也很简单。能够将命令模块转变为模拟器显然是可取的——有许多可用的命令模块。我向训练中心的指挥官提出了建议,并获得了进行演示的许可。我去了台北的电子元件商店,买了一些零件。我花了大约一天的时间将所有东西组合在一起,然后我能够将自己的视频信号注入雷达屏幕上。指挥官很兴奋,并要求我写一份正式的提案。最终被拒绝了,因为我绝对无意在正常服务之外继续留下来。

The Missile Command had a few simulators for training officers in charge of firing the missiles. The simulators were expensive, and they were all broken. Before I became an instructor, I went through maintenance training for the command module, and was somewhat familiar with its internal structure. The simulator was based on a modified command module, and was effectively a glorified video game that used the radar screen as the display. The Center had the schematics for both the simulator and the command module. They were based on ancient technologies: relay-based logic, analog potentiometers for waveform generation, and so on. Typical military stuff, ten to twenty years out of date. The video stream for the command module was easily accessible, and modifying it was trivial. To be able to turn a command module into a simulator was clearly desirable—there were many working command modules. I made a proposal to the Commander of the Training Center, and got the go-ahead to do a demonstration. I went to the electronic-component stores in Taipei and acquired a few parts. It took about a day to put things together, and I was able to inject my own video signal onto the radar screen. The Commander was thrilled and I was asked to write a formal proposal. It was turned down in the end, as I had absolutely no intention of staying on beyond my regular service.

服兵役对我来说是一次痛苦的经历。从身体上来说,我很好。心理上,长期的精神压抑也造成了损害。我的心智完全恢复之前,大概是服兵役后快两年了。在军队期间,我设法做了一些不同寻常的事情,但两年无条件的权威对我来说实在是太难以承受了。回想起来,我想知道我在研究生院早期的越轨行为是否是我在服兵役期间长期被剥夺自由意志的结果。

Military service was a traumatic experience for me. Physically, I was fine. Psychologically, the long period of mental suppression took its toll. It was probably close to two years after military service before my mental acumen fully recovered. I managed to do a few things out of the ordinary while in the military, but being subject to two years of unconditional authority was simply too much for me. Looking back, I wonder whether my early escapades in graduate school were a result of long deprivation of free will during my military service.

我服兵役的最后一年是 1982 年。那年早些时候,我收到卡内基梅隆大学的通知,我被他们的计算机科学系录取。因此,那年八月,退伍几周后,我登上飞机前往美国。

My last year of military service was 1982. Early in that year I received notification from Carnegie Mellon University that I was accepted into their Computer Science Department. So, that August, a couple of weeks after leaving military service, I got on the plane and left for the United States.

卡内基梅隆大学的案例

The Case for Carnegie Mellon

我是怎么来到卡内基梅隆大学的?嗯,这有点复杂。

How did I end up at Carnegie Mellon? Well, it was kind of convoluted.

当我在国立台湾大学电机工程系读本科时,台湾的信息技术(IT)产业几乎不存在。微处理器革命才刚刚开始。看到微处理器可能为台湾IT产业的跨越提供一条途径,台湾国科会资助​​大学的微处理器研究项目。结果,在大二的时候,我开始为我们的部门主管从事微处理器研究项目。

The Taiwanese information technology (IT) industry was virtually nonexistent when I was an undergraduate in the Electrical Engineering (EE) Department of National Taiwan University. The microprocessor revolution had just begun. Seeing that the microprocessor might provide a way for the Taiwanese IT industry to leapfrog, the National Science Council in Taiwan funded microprocessor research projects in universities. As a result, in my sophomore year I began to work for our department head on a microprocessor research project.

在我的同学 Chia-Min Hwang 问我是否有兴趣与他合作后,我加入了这个项目关于该项目。对于这个项目,与国际象棋项目中发生的情况相反,我是软件人员,黄是硬件人员。我们的项目是对机床进行振动分析。这个项目一直持续到我毕业。回想起来,我的工作并没有太多真正的研究,但我培养了一些有用的计算机技能。这段经历对我选择研究生院起到了一定的作用。

I joined the project after a classmate of mine, Chia-Min Hwang, asked whether I was interested in partnering him to work on the project. For this project, opposite to what happened in the chess project, I was the software guy and Hwang was the hardware guy. Our project was to do vibration analysis of tooling machines. The project lasted until my graduation. In retrospect, my work was not much in terms of real research, but I developed some useful computer skills. This experience played a role in my choice of graduate school.

大学的微处理器研究项目几乎都是软件项目。IT 硬件的研究实际上超出了台湾大学的能力范围。为了处理硬件研究,政府设立了工业研究机构,其中之一是电子研究和服务组织(ERSO)。在我大四的时候,ERSO 的一位代表给电子工程系的学生做了一次演讲。他强调,硅芯片对我们生活的方方面面都至关重要,鉴于台湾自然资源匮乏,发展IC产业对未来经济增长至关重要。ERSO 的演讲者 Chintay Shih 现任工业技术研究院院长,该研究院包括 ERSO 本身。这次演讲一定很精彩,因为后来我决定要研究如何设计芯片7

The university-based microprocessor research projects were almost all software projects. Research in IT hardware was effectively out of the reach of Taiwanese universities. To handle hardware research, the government set up industrial research agencies, one of which was the Electronics Research and Services Organization (ERSO). In my senior year, a representative from ERSO gave a talk to students in the EE Department. He stressed that silicon chips would be critical in every aspect of our lives and, given the lack of natural resources in Taiwan, developing an IC industry would be vital for the future growth of the economy. The speaker from ERSO, Chintay Shih, is today the President of the Industrial Technology Research Institute, which includes ERSO itself. The talk must have been a good one, because afterwards I decided that I wanted to study how to design chips7.

当时学习芯片设计最好的地方是美国,所以我向美国的大学发出了几份申请。我的大部分申请都发送给了 EE 部门。不过,我确实向匹兹堡卡内基梅隆大学计算机科学 (CS) 系发送了一份。我在那里发送申请有两个原因。

At the time the best place to study chip design was the United States, so I sent several applications to US universities. Most of my applications were sent to EE Departments. However, I did send one to the Computer Science (CS) Department at Carnegie Mellon University in Pittsburgh. There were two reasons that I sent my application there.

第一个原因是,只要我从事芯片设计工作,我的微处理器经验使我可以接受计算机科学系的想法。卡内基梅隆大学的计算机科学系排名前三,这也没有什么坏处。

The first reason was that my microprocessor experience made the idea of a CS department palatable, as long as I got to work on chip design. The fact that the CS Department at Carnegie Mellon was ranked among the top three did not hurt either.

第二个原因与我在卡内基梅隆大学的导师 HT Kung 教授有关,但不是你想象的那样。我选择卡内基梅隆大学并不是因为那里有孔教授,而是因为那里有孔教授。情况有点复杂。20世纪70年代末和80年代初,孔教授竭尽全力帮助台湾新兴的IC产业站稳脚跟。中的一个他所做的就是向台湾研究人员介绍 Carver Mead 和 Lynn Conway 所著的《VLSI 系统导论》一书。这本书是 20 世纪 80 年代初 VLSI 课程的圣经。书中有一章由Charles Leiserson(当时孔教授的学生)和孔教授本人撰写。顺便说一句,查尔斯现在是麻省理工学院的教授。无论如何,从本章可以明显看出,卡内基梅隆大学计算机科学系有一个活跃且受人尊敬的 VLSI 小组。这是我提交申请的第二个原因。

The second reason had something to do with my faculty advisor at Carnegie Mellon, Professor H T Kung, but not in the way you might think. I did not choose Carnegie Mellon because of Professor Kung’s presence there; it was a bit more complicated. In the late 1970s and early 1980s, Professor Kung did what he could to help the budding IC industry in Taiwan to get on its feet. One of the things that he did was to introduce the book Introduction to VLSI Systems by Carver Mead and Lynn Conway, to Taiwanese researchers. This book was the bible for VLSI classes in the early 1980s. In the book, there was a chapter by Charles Leiserson (then a student of Professor Kung’s) and Professor Kung himself. Charles, by the way, is now a professor at MIT. Anyway, from this chapter it was obvious that there was an active and well-respected VLSI group at the CS Department of Carnegie Mellon. This was the second reason for my sending in the application.

我收到了两份我真正感兴趣的学校的录取通知书。其中一份是位于硅谷中心的斯坦福大学电子工程系。另一件是卡内基梅隆大学计算机科学系的。凭借我的电气工程背景,斯坦福大学本来是我的第一选择,但卡内基梅隆大学提供的研究助理职位改变了我的选择,所以我去了匹兹堡

I received two admissions from schools that I was really interested in. One was for the EE Department at Stanford University, in the heart of the Silicon Valley. The other one was for the CS Department at Carnegie Mellon University. With my electrical engineering background, Stanford would have been my first choice, but the research assistantship offered by Carnegie Mellon tipped the scale, and I went to Pittsburgh

    事实上,即使你只是一位长相普通的计算机科学家,而对手是一位非常优秀的绅士,它们也能很好地发挥作用。

    Actually, they work well even if you are only an ordinary looking computer scientist and the opponent is a perfectly fine gentleman.

    达到第 8 级的棋子将晋升为更强大的棋子,通常是皇后,是游戏中最强大的棋子。

    A pawn reaching the 8th rank is promoted to a more powerful piece, usually a queen, the most powerful piece in the game.

    Go 中的排名系统是一个三层系统。在每个级别中,玩家按照级数数。在最低层,玩家按照kyu的数量进行排名,实力越强的kyu越少。一阶玩家是最强的九级玩家。九级以上,有业余段位选手,也有职业段位选手。在段位上,段位越高的选手越好。一段选手,也称为shodans,是最弱的段位选手。无论是业余段位还是职业段位,排名最高的选手都是九段。业余九段选手通常并不比职业首段选手强。业余三段棋手大致相当于国际象棋大师级棋手。

    The ranking system in Go is a three-tiered system. Within each tier, players are ranked either by the number of kyu’s or the number of dan’s they have. At the lowest tier, the players are ranked by the number of kyu’s, with stronger players having fewer kyu’s. The one-kyu players are the strongest kyu-level players. Above the kyu-level, we have the amateur dan-level players, and the professional dan-level players. At dan-level, a player with more dan’s is the better player. One-dan players, also known as shodans, are the weakest dan-level players. For both amateur and professional dan levels, the highest ranked players are 9-dans. The amateur 9-dans are usually no stronger than professional shodans. An amateur 3-dan player roughly corresponds to a master level player in chess.

    正国据说是台湾最好的公立男高中,入学难度很大,因此有入学礼物。

    Cheng-Kuo is said to be the best public high school for boys in Taiwan and quite difficult to get in. Hence the gift for achieving entrance.

    如果您不知道什么是有限差分方程,请不要担心。我发现这些知识很有用,但你当然可以在不知道它是什么的情况下阅读本书的其余部分。

    If you don’t know what a finite difference equation is, don’t worry about it. I found the knowledge useful, but you certainly can read the rest of the book without knowing what it is.

    在中文里,“疯”这个词与“峰”同音,“峰”字是“峰”字,也是我名字的第一个字。

    In Chinese, the word “crazy” has the same sound as “feng”, the Chinese character for “peak” and the first character of my given name.

    好吧,我有点夸张了。我对学习集成电路的工作原理很感兴趣。我的一个个人项目迫使我了解这个主题。

    Well, I am exaggerating a little bit. I had an existing interest in learning how integrated circuits work. A personal project of mine forced me to learn about the subject.

 

 

附录B

APPENDIX B

选定的比赛分数

Selected Game Scores


Chiptest-Recom,1986 年 ACM 锦标赛,达拉斯,第 5 轮

Chiptest-Recom, 1986 ACM Championship, Dallas, Round 5

1. f4 d5 2. e3 c5 3. Bb5+ Bd7 4. Be2 g6 5. Nf3 Bg7 6. c4 dxc4 7. Bxc4 b5 8. Be2 Bc6 9. OO c4 10. a4 a6 11. axb5 axb5 12. Rxa8 Bxa8 13​​. Na3 Bxf3 14.Rxf3 Qd5 15.b3 cxb3 16.Bxb5+ Nd7 17.Bc4 Qb7 18.Bxb3 Ngf6 19.Rf1 OO 20.Nc4 Ra8 21.Ne5 e6 22.Nd3 Ne4 23.Nf2 Nxf2 24.Kxf2 Nc5 25. BC4钠26. Qc2 Nb6 27. Bb2 Nxc4 28. Bxg7 Nxd2 29. Qxd2 Kxg7 30. Kg1 Ra4 31. Qc3+ f6 32. Rc1 Qe4 33. Qb3 Ra7 34. Rc8 e5 35. Rc4 Qa8 36. Kf2 exf4 37. exf4 Ra2+ 38. Rc2 Qa7+ 39.Kf1 Ral+ 40.Ke2 Qe7+ 41.Kd2 Qd6+ 42.Ke2 Qxf4 43.h3 Qe4+ 44.Kf2 Qh4+ 45.g3 Qxh3 46.Rc7+ Kh6 47.Qe3+ g5 48.Rxh7+ Kxh7 49.Qe7+ Kg 6 1/2- 1/2。

1. f4 d5 2. e3 c5 3. Bb5+ Bd7 4. Be2 g6 5. Nf3 Bg7 6. c4 dxc4 7. Bxc4 b5 8. Be2 Bc6 9. O-O c4 10. a4 a6 11. axb5 axb5 12. Rxa8 Bxa8 13. Na3 Bxf3 14. Rxf3 Qd5 15. b3 cxb3 16. Bxb5+ Nd7 17. Bc4 Qb7 18. Bxb3 Ngf6 19. Rf1 O-O 20. Nc4 Ra8 21. Ne5 e6 22. Nd3 Ne4 23. Nf2 Nxf2 24. Kxf2 Nc5 25. Bc4 Na4 26. Qc2 Nb6 27. Bb2 Nxc4 28. Bxg7 Nxd2 29. Qxd2 Kxg7 30. Kg1 Ra4 31. Qc3+ f6 32. Rc1 Qe4 33. Qb3 Ra7 34. Rc8 e5 35. Rc4 Qa8 36. Kf2 exf4 37. exf4 Ra2+ 38. Rc2 Qa7+ 39. Kf1 Ral+ 40. Ke2 Qe7+ 41. Kd2 Qd6+ 42. Ke2 Qxf4 43. h3 Qe4+ 44. Kf2 Qh4+ 45. g3 Qxh3 46. Rc7+ Kh6 47. Qe3+ g5 48. Rxh7+ Kxh7 49. Qe7+ Kg6 1/2-1/2.

Bent Larsen-Deep Thought,1988 年 Software Toolworks 国际象棋锦标赛,长滩,第 3 轮

Bent Larsen-Deep Thought, 1988 Software Toolworks Chess Championship, Long Beach, Round 3

1. c4 e5 2. g3 Nf6 3. Bg2 c6 4. Nf3 e4 5. Nd4 d5 6. cxd5 Qxd5 7. Nc2 Qh5 8. h4 Bf5 9. Ne3 Bc5 10. Qb3 b6 11. Qa4 OO 12. Nc3 b5 13. Qc2 Bxe3 14. dxe3 Re8 15. a4 b4 16. Nb1 Nbd7 17. Nd2 Re6?(Ne5!,想法是 Nf3 放弃 c 兵以进行强力攻击,稍后可能会使用 Bd3) 18. b3 Rd8 19. Bb2 Bg6 20. Nc4 Nd5 21. OOO N7f6 22. Bh3 Bf5 23. Bxf5 Qxf5 24. f3 h5 25.Bd4 Rd7 26.Kb2 Rc7 27.g4 hxg4 28.Rhg1 c5 29.fxg4 Nxg4 30.Bxg7 Rg6 31.Qd2 Rd7 32.Rxg4 Rxg4 33.Ne5 Nxe3 34.Qxd7 Nxd1 35.Qxd1 Rg3 36.Qd6 Kxg7 37. Nd7 Re3 38. Qh2 Kh7 39. Nf8 Kh8 40. h5 Qd5 41. Ng6 fxg6 42. hxg6 Kg7 43. Qh7 Kf6 0-1

1. c4 e5 2. g3 Nf6 3. Bg2 c6 4. Nf3 e4 5. Nd4 d5 6. cxd5 Qxd5 7. Nc2 Qh5 8. h4 Bf5 9. Ne3 Bc5 10. Qb3 b6 11. Qa4 O-O 12. Nc3 b5 13. Qc2 Bxe3 14. dxe3 Re8 15. a4 b4 16. Nb1 Nbd7 17. Nd2 Re6? (Ne5!, with the idea of Nf3 giving up the c pawn for a strong attack, possibly with Bd3 later) 18. b3 Rd8 19. Bb2 Bg6 20. Nc4 Nd5 21. O-O-O N7f6 22. Bh3 Bf5 23. Bxf5 Qxf5 24. f3 h5 25. Bd4 Rd7 26. Kb2 Rc7 27. g4 hxg4 28. Rhg1 c5 29. fxg4 Nxg4 30. Bxg7 Rg6 31. Qd2 Rd7 32. Rxg4 Rxg4 33. Ne5 Nxe3 34. Qxd7 Nxd1 35. Qxd1 Rg3 36. Qd6 Kxg7 37. Nd7 Re3 38. Qh2 Kh7 39. Nf8 Kh8 40. h5 Qd5 41. Ng6 fxg6 42. hxg6 Kg7 43. Qh7 Kf6 0-1

深蓝-加里·卡斯帕罗夫,1996 年 ACM 国际象棋挑战赛,费城,第 1 场

Deep Blue-Garry Kasparov, 1996 ACM Chess Challenge, Philadelphia, Game 1

1.e4 c5 2.c3 d5 3.exd5 Qxd5 4.d4 Nf6 5.Nf3 Bg4 6.Be2 e6 7.h3 Bh5 8.0-0 Nc6 9.Be3 cxd4 10.cxd4 Bb4 11.a3 Ba5 12.Nc3 Qd6 13. Nb5 Qe7 14.Ne5 Bxe2 15.Qxe2 0-0 16.Rac1 Rac8 17.Bg5 Bb6 18.Bxf6 gxf6 19.Nc4 Rfd8 20.Nxb6 axb6 21.Rfd1 f5 22.Qe3 Qf6 23.d5 Rxd5 24.Rxd5 前d5 25。 b3 Kh8 26.Qxb6 Rg8 27.Qc5 d4 28.Nd6 f4 29.Nxb7 Ne5 30.Qd5 f3 31.g3 Nd3 32.Rc7 Re8 33.Nd6 Rel+ 34.Kh2 Nxf2 35.Nxf7+ Kg7 36.Ng5+ Kh6 37.Rxh7 + 1 -0

1.e4 c5 2.c3 d5 3.exd5 Qxd5 4.d4 Nf6 5.Nf3 Bg4 6.Be2 e6 7.h3 Bh5 8.0-0 Nc6 9.Be3 cxd4 10.cxd4 Bb4 11.a3 Ba5 12.Nc3 Qd6 13.Nb5 Qe7 14.Ne5 Bxe2 15.Qxe2 0-0 16.Rac1 Rac8 17.Bg5 Bb6 18.Bxf6 gxf6 19.Nc4 Rfd8 20.Nxb6 axb6 21.Rfd1 f5 22.Qe3 Qf6 23.d5 Rxd5 24.Rxd5 exd5 25.b3 Kh8 26.Qxb6 Rg8 27.Qc5 d4 28.Nd6 f4 29.Nxb7 Ne5 30.Qd5 f3 31.g3 Nd3 32.Rc7 Re8 33.Nd6 Rel+ 34.Kh2 Nxf2 35.Nxf7+ Kg7 36.Ng5+ Kh6 37.Rxh7+ 1-0

Garry Kasparov-深蓝,1996 年 ACM 国际象棋挑战赛,费城,第 2 场

Garry Kasparov-Deep Blue, 1996 ACM Chess Challenge, Philadelphia, Game 2

1.Nf3 d5 2.d4 e6 3.g3 c5 4.Bg2 Nc6 5.0-0 Nf6 6.c4 dxc4 7.Ne5 Bd7 8.Na3 cxd4 9.Naxc4 Bc5 10.Qb3 0-0 11.Qxb7 Nxe5 12.Nxe5 Rb8 13.Qf3 Bd6 14.NC6 Bxc6 15.Qxc6 e5 16.Rb1 Rb6 17.Qa4 Qb8 18.Bg5 Be7 19.b4 Bxb4 20.Bxf6 gxf6 21.Qd7 Qc8 22.Qxa7 Rb8 23.Qa4 Bc3 24.Rxb8 Q xb8 25。 Be4 Qc7 26.Qa6 Kg7 27.Qd3 Rb8 28.Bxh7 Rb2 29.Be4 Rxa2 30.h4 Qc8 31.Qf3 Ra1 32.Rxal Bxa1 33.Qh5 Qh8 34.Qg4+ Kf8 35.Qc8+ Kg7 36.Qg4+ Kf8 37. Bd5 Ke7 38.Bc6 Kf8 39.Bd5 Ke7 40.Qf3 Bc3 41.Bc4 Qc8 42.Qd5 Qe6 43.Qb5 Qd7 44.Qc5+ Qd6 45.Qa7+ Qd7 46.Qa8 Qc7 47.Qa3+ Qd6 48.Qa2 f5 49.Bxf7 e4 50. Bh5 Qf6 51.Qa3+ Kd7 52.Qa7+ Kd8 53.Qb8+ Kd7 54.Be8+ Ke7 55.Bb5 Bd2 56.Qc7+ Kf8 57.Bc4 Bc3 58.Kg2 Bel 59.Kf1 Bc3 60.f4 exf3 61.exf3 Bd2 62.f4柯8 63.QC8+ KE7 64.QC5+ KD8 65.BD3 BE3 66.QXF5 QC6 67.QF8+ KC7 68.QE7+ KC8 69.BF5+ KB8 70.QD8+ KB7 71.QD7 71.QD7+ QXD7 72.BXD7 72.BXD7 KC7 KC7 KC7 KC7 73.BBBBB56 1-0-0-0-0

1.Nf3 d5 2.d4 e6 3.g3 c5 4.Bg2 Nc6 5.0-0 Nf6 6.c4 dxc4 7.Ne5 Bd7 8.Na3 cxd4 9.Naxc4 Bc5 10.Qb3 0-0 11.Qxb7 Nxe5 12.Nxe5 Rb8 13.Qf3 Bd6 14.Nc6 Bxc6 15.Qxc6 e5 16.Rb1 Rb6 17.Qa4 Qb8 18.Bg5 Be7 19.b4 Bxb4 20.Bxf6 gxf6 21.Qd7 Qc8 22.Qxa7 Rb8 23.Qa4 Bc3 24.Rxb8 Qxb8 25.Be4 Qc7 26.Qa6 Kg7 27.Qd3 Rb8 28.Bxh7 Rb2 29.Be4 Rxa2 30.h4 Qc8 31.Qf3 Ra1 32.Rxal Bxa1 33.Qh5 Qh8 34.Qg4+ Kf8 35.Qc8+ Kg7 36.Qg4+ Kf8 37.Bd5 Ke7 38.Bc6 Kf8 39.Bd5 Ke7 40.Qf3 Bc3 41.Bc4 Qc8 42.Qd5 Qe6 43.Qb5 Qd7 44.Qc5+ Qd6 45.Qa7+ Qd7 46.Qa8 Qc7 47.Qa3+ Qd6 48.Qa2 f5 49.Bxf7 e4 50.Bh5 Qf6 51.Qa3+ Kd7 52.Qa7+ Kd8 53.Qb8+ Kd7 54.Be8+ Ke7 55.Bb5 Bd2 56.Qc7+ Kf8 57.Bc4 Bc3 58.Kg2 Bel 59.Kf1 Bc3 60.f4 exf3 61.exf3 Bd2 62.f4 Ke8 63.Qc8+ Ke7 64.Qc5+ Kd8 65.Bd3 Be3 66.Qxf5 Qc6 67.Qf8+ Kc7 68.Qe7+ Kc8 69.Bf5+ Kb8 70.Qd8+ Kb7 71.Qd7+ Qxd7 72.Bxd7 Kc7 73.Bb5 Kd6 1-0

深蓝-加里·卡斯帕罗夫,1996 年 ACM 国际象棋挑战赛,费城,第 3 场

Deep Blue-Garry Kasparov, 1996 ACM Chess Challenge, Philadelphia, Game 3

1.e4 c5 2.c3 d5 3.exd5 Qxd5 4.d4 Nf6 5.Nf3 Bg4 6.Be2 e6 7.0-0 Nc6 8.Be3 cxd4 9.cxd4 Bb4 10.a3 Ba5 11.Nc3 Qd6 12.Ne5 Bxe2 13. Qxe2 Bxc3 14.bxc3 Nxe5 15.Bf4 Nf3+ 16.Qxf3 Qd5 17.Qd3 Rc8 18.Rfc1 Qc4 19.Qxc4 Rxc4 20.Rcb1 b6 21.Bb8 Ra4 22.Rb4 Ra5 23.Rc4 0-0 24.Bd6 Ra 8月25日。 RC6 B5 26.KF1 RA4 27.RB1 A6 28.KE2 H5 29.KD3 RD8 30.BE7 RD7 RD7 31.BXF6 GXF6 GXF6 GXF6 32.RB3 KG7 33.KE3 E5 34.G3 EXD4 EXD4 exd4+ 35.cxd4+ 35.cxd4 re7+ 36.kf3 rd7 37.kf3 rd7 37.raxd3 Raxd3 Raxd3 Raxd4 Raxd4 Raxd4 Raxd4 38.Rxd4 Rxd4 39.Rxa6 1/2-1/2

1.e4 c5 2.c3 d5 3.exd5 Qxd5 4.d4 Nf6 5.Nf3 Bg4 6.Be2 e6 7.0-0 Nc6 8.Be3 cxd4 9.cxd4 Bb4 10.a3 Ba5 11.Nc3 Qd6 12.Ne5 Bxe2 13.Qxe2 Bxc3 14.bxc3 Nxe5 15.Bf4 Nf3+ 16.Qxf3 Qd5 17.Qd3 Rc8 18.Rfc1 Qc4 19.Qxc4 Rxc4 20.Rcb1 b6 21.Bb8 Ra4 22.Rb4 Ra5 23.Rc4 0-0 24.Bd6 Ra8 25.Rc6 b5 26.Kf1 Ra4 27.Rb1 a6 28.Ke2 h5 29.Kd3 Rd8 30.Be7 Rd7 31.Bxf6 gxf6 32.Rb3 Kg7 33.Ke3 e5 34.g3 exd4+ 35.cxd4 Re7+ 36.Kf3 Rd7 37.Rd3 Raxd4 38.Rxd4 Rxd4 39.Rxa6 1/2-1/2

Garry Kasparov-深蓝,1996 年 ACM 国际象棋挑战赛,费城,第 4 场

Garry Kasparov-Deep Blue, 1996 ACM Chess Challenge, Philadelphia, Game 4

1.Nf3 d5 2.d4 c6 3.c4 e6 4.Nbd2 Nf6 5.e3 Nbd7 6.Bd3 Bd6 7.e4 dxe4 8.Nxe4 Nxe4 9.Bxe4 0-0 10.0-0 h6 11.Bc2 e5 12.Rel exd4 13.Qxd4 Bc5 14.Qc3 a5 15.a3 Nf6 16.Be3 Bxe3 17.Rxe3 Bg4 18.Ne5 Re8 19.Rae1 Be6 20.f4 Qc8 21.h3 b5 22.f5 Bxc4 23.Nxc4 bxc4 24.Rxe8+ Nxe8 25 。 Re4 Nf6 26.Rxc4 Nd5 27.Qe5 Qd7 28.Rg4 f6 29.Qd4 Kh7 30.Re4 Rd8 31.Khl Qc7 32.Qf2 Qb8 33.Ba4 c5 34.Bc6 c4 35.Rxc4 Nb4 36.Bf3 Nd3 37.Qh4 Qxb2 38.Qg3 Qxa3 39.Rc7 Qf8 40.Ra7 Ne5 41.Rxa5 Qf7 42.Rxe5 fxe5 43.Qxe5 Re8 44.Qf4 Qf6 45.Bh5 Rf8 46.Bg6+ Kh8 47.Qc7 Qd4 48.Kh2 Ra8 49.Bh5 Qf6 50. Bg6 Rg8 1/2-1/2

1.Nf3 d5 2.d4 c6 3.c4 e6 4.Nbd2 Nf6 5.e3 Nbd7 6.Bd3 Bd6 7.e4 dxe4 8.Nxe4 Nxe4 9.Bxe4 0-0 10.0-0 h6 11.Bc2 e5 12.Rel exd4 13.Qxd4 Bc5 14.Qc3 a5 15.a3 Nf6 16.Be3 Bxe3 17.Rxe3 Bg4 18.Ne5 Re8 19.Rae1 Be6 20.f4 Qc8 21.h3 b5 22.f5 Bxc4 23.Nxc4 bxc4 24.Rxe8+ Nxe8 25.Re4 Nf6 26.Rxc4 Nd5 27.Qe5 Qd7 28.Rg4 f6 29.Qd4 Kh7 30.Re4 Rd8 31.Khl Qc7 32.Qf2 Qb8 33.Ba4 c5 34.Bc6 c4 35.Rxc4 Nb4 36.Bf3 Nd3 37.Qh4 Qxb2 38.Qg3 Qxa3 39.Rc7 Qf8 40.Ra7 Ne5 41.Rxa5 Qf7 42.Rxe5 fxe5 43.Qxe5 Re8 44.Qf4 Qf6 45.Bh5 Rf8 46.Bg6+ Kh8 47.Qc7 Qd4 48.Kh2 Ra8 49.Bh5 Qf6 50.Bg6 Rg8 1/2-1/2

深蓝-加里·卡斯帕罗夫,1996 年 ACM 国际象棋挑战赛,费城,第 5 场

Deep Blue-Garry Kasparov, 1996 ACM Chess Challenge, Philadelphia, Game 5

1.e4 e5 2.Nf3 Nf6 3.Nc3 Nc6 4.d4 exd4 5.Nxd4 Bb4 6.Nxc6 bxc6 7.Bd3 d5 8.exd5 cxd5 9.0-0 0-0 10.Bg5 c6 11.Qf3 Be7 12.Rael Re8 13.Ne2 h6 14.Bf4 Bd6 15.Nd4 Bg4 16.Qg3 Bxf4 17.Qxf4 Qb6 18.c4 Bd7 19.cxd5 cxd5 20.Rxe8+ Rxe8 21.Qd2 Ne4 22.Bxe4 dxe4 23.b3 Rd8 24.Qc3 f5 25。 RD1 BE6 26.QE3 BF7 27.QC3 F4 28.RD2 QF6 29.G3 RD5 30.A3 KH7 31.KG2 QE5 QE5 32.F3 E3 33.RD3 E2 E2 34.GXF4 E1Q 35.FXE5 QXC3 36.RXC3 36.RXC3 RXC3 RXC3 RXD4 37.BB4 37.BB4 37.BC4 BC4 38.Kf2 g5 39.Re3 Rd2+ 40.Kel Rd3 41.Kf2 Kg6 42.Rxd3 Bxd3 43.Ke3 Bc2 44.Kd4 Kf5 45.Kd5 h5 0-1

1.e4 e5 2.Nf3 Nf6 3.Nc3 Nc6 4.d4 exd4 5.Nxd4 Bb4 6.Nxc6 bxc6 7.Bd3 d5 8.exd5 cxd5 9.0-0 0-0 10.Bg5 c6 11.Qf3 Be7 12.Rael Re8 13.Ne2 h6 14.Bf4 Bd6 15.Nd4 Bg4 16.Qg3 Bxf4 17.Qxf4 Qb6 18.c4 Bd7 19.cxd5 cxd5 20.Rxe8+ Rxe8 21.Qd2 Ne4 22.Bxe4 dxe4 23.b3 Rd8 24.Qc3 f5 25.Rd1 Be6 26.Qe3 Bf7 27.Qc3 f4 28.Rd2 Qf6 29.g3 Rd5 30.a3 Kh7 31.Kg2 Qe5 32.f3 e3 33.Rd3 e2 34.gxf4 e1Q 35.fxe5 Qxc3 36.Rxc3 Rxd4 37.b4 Bc4 38.Kf2 g5 39.Re3 Rd2+ 40.Kel Rd3 41.Kf2 Kg6 42.Rxd3 Bxd3 43.Ke3 Bc2 44.Kd4 Kf5 45.Kd5 h5 0-1

Garry Kasparov-深蓝,1996 年 ACM 国际象棋挑战赛,费城,第 6 场

Garry Kasparov-Deep Blue, 1996 ACM Chess Challenge, Philadelphia, Game 6

1.Nf3 d5 2.d4 c6 3.c4 e6 4.Nbd2 Nf6 5.e3 c5 6.b3 Nc6 7.Bb2 cxd4 8.exd4 Be7 9.Rc1 0-0 10.Bd3 Bd7 11.0-0 Nh5 12.Rel Nf4 13.Bb1 Bd6 14.g3 Ng6 15.Ne5 Rc8 16.Nxd7 Qxd7 17.Nf3 Bb4 18.Re3 Rfd8 19.h4 Nge7 20.a3 Ba5 21.b4 Bc7 22.c5 Re8 23.Qd3 g6 24.Re2 Nf5 25. Bc3 h5 26.b5 Nce7 27.Bd2 Kg7 28.a4 Ra8 29.a5 a6 30.b6 Bb8 31.Bc2 Nc6 32.Ba4 Re7 33.Bc3 Ne5 34.dxe5 Qxa4 35.Nd4 Nxd4 36.Qxd4 Qd7 37.Bd2 Re8 38.Bg5 Rc8 39.Bf6+ Kh7 40.c6 bxc6 41.Qc5 Kh6 42.Rb2 Qb7 43.Rb4 1-0

1.Nf3 d5 2.d4 c6 3.c4 e6 4.Nbd2 Nf6 5.e3 c5 6.b3 Nc6 7.Bb2 cxd4 8.exd4 Be7 9.Rc1 0-0 10.Bd3 Bd7 11.0-0 Nh5 12.Rel Nf4 13.Bb1 Bd6 14.g3 Ng6 15.Ne5 Rc8 16.Nxd7 Qxd7 17.Nf3 Bb4 18.Re3 Rfd8 19.h4 Nge7 20.a3 Ba5 21.b4 Bc7 22.c5 Re8 23.Qd3 g6 24.Re2 Nf5 25.Bc3 h5 26.b5 Nce7 27.Bd2 Kg7 28.a4 Ra8 29.a5 a6 30.b6 Bb8 31.Bc2 Nc6 32.Ba4 Re7 33.Bc3 Ne5 34.dxe5 Qxa4 35.Nd4 Nxd4 36.Qxd4 Qd7 37.Bd2 Re8 38.Bg5 Rc8 39.Bf6+ Kh7 40.c6 bxc6 41.Qc5 Kh6 42.Rb2 Qb7 43.Rb4 1-0

Garry Kasparov-深蓝,IBM 卡斯帕罗夫 vs. 深蓝复赛,第 1 场

Garry Kasparov-Deep Blue, IBM Kasparov vs. Deep Blue Rematch, Game 1

1.Nf3 d5 2.g3 Bg4 3.b3 Nd7 4.Bb2 e6 5.Bg2 Ngf6 6.OO c6 7.d3 Bd6 8.Nbd2 OO 9.h3 Bh5 10.e3 h6 11.Qe1 Qa5 12.a3 Bc7 13. Nh4 g5 14.Nhf3 e5 15.e4 Rfe8 16.Nh2 Qb6 17.Qc1 a5 18.Rel Bd6 19.Ndf1 dxe4 20.dxe4 Bc5 21.Ne3 Rad8 22.Nhf1 g4 23.hxg4 Nxg4 24.f3 Nxe3 25.N xe3 Be7 26.Khl Bg5 27.Re2 a4 28.b4 f5 29.exf5 e4 30.f4 Bxe2 31.fxg5 Ne5 32.g6 Bf3 33.Bc3 Qb5 34.Qf1 Qxf1+ 35.Rxf1 h5 36.Kg1 Kf8 37.Bh3 b5 38. Kf2 Kg7 39.g4 Kh6 40.Rg1 hxg4 41.Bxg4 Bxg4 42.Nxg4+ Nxg4+ 43.Rxg4 Rd5 44.f6 Rd1 45.g7 1-0

1.Nf3 d5 2.g3 Bg4 3.b3 Nd7 4.Bb2 e6 5.Bg2 Ngf6 6.O-O c6 7.d3 Bd6 8.Nbd2 O-O 9.h3 Bh5 10.e3 h6 11.Qe1 Qa5 12.a3 Bc7 13.Nh4 g5 14.Nhf3 e5 15.e4 Rfe8 16.Nh2 Qb6 17.Qc1 a5 18.Rel Bd6 19.Ndf1 dxe4 20.dxe4 Bc5 21.Ne3 Rad8 22.Nhf1 g4 23.hxg4 Nxg4 24.f3 Nxe3 25.Nxe3 Be7 26.Khl Bg5 27.Re2 a4 28.b4 f5 29.exf5 e4 30.f4 Bxe2 31.fxg5 Ne5 32.g6 Bf3 33.Bc3 Qb5 34.Qf1 Qxf1+ 35.Rxf1 h5 36.Kg1 Kf8 37.Bh3 b5 38.Kf2 Kg7 39.g4 Kh6 40.Rg1 hxg4 41.Bxg4 Bxg4 42.Nxg4+ Nxg4+ 43.Rxg4 Rd5 44.f6 Rd1 45.g7 1-0

深蓝-Garry Kasparov,IBM 卡斯帕罗夫 vs. 深蓝复赛,第 2 场

Deep Blue-Garry Kasparov, IBM Kasparov vs. Deep Blue Rematch, Game 2

1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.OO Be7 6.Rel b5 7.Bb3 d6 8.c3 OO 9.h3 h6 10.d4 Re8 11.Nbd2 Bf8 12.Nf1 Bd7 13. Ng3 Na5 14.Bc2 c5 15.b3 Nc6 16.d5 Ne7 17.Be3 Ng6 18.Qd2 Nh7 19.a4 Nh4 20.Nxh4 Qxh4 21.Qe2 Qd8 22.b4 Qc7 23.Rec1 c4 24.Ra3 Rec8 25.Rca1 Qd8 26.f4 Nf6 27.fxe5 dxe5 28.Qf1 Ne8 29.Qf2 Nd6 30.Bb6 Qe8 31.R3a2 Be7 32.Bc5 Bf8 33.Nf5 Bxf5 34.exf5 f6 35.Bxd6 Bxd6 36.axb5 axb5 37.Be4 Rx a2 38. Qxa2 Qd7 39.Qa7 Rc7 40.Qb6 Rb7 41.Ra8+ Kf7 42.Qa6 Qc7 43.Qc6 Qb6+ 44.Kf1 Rb8 45.Ra6 1-0

1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 5.O-O Be7 6.Rel b5 7.Bb3 d6 8.c3 O-O 9.h3 h6 10.d4 Re8 11.Nbd2 Bf8 12.Nf1 Bd7 13.Ng3 Na5 14.Bc2 c5 15.b3 Nc6 16.d5 Ne7 17.Be3 Ng6 18.Qd2 Nh7 19.a4 Nh4 20.Nxh4 Qxh4 21.Qe2 Qd8 22.b4 Qc7 23.Rec1 c4 24.Ra3 Rec8 25. Rca1 Qd8 26.f4 Nf6 27.fxe5 dxe5 28.Qf1 Ne8 29.Qf2 Nd6 30.Bb6 Qe8 31.R3a2 Be7 32.Bc5 Bf8 33.Nf5 Bxf5 34.exf5 f6 35.Bxd6 Bxd6 36.axb5 axb5 37.Be4 Rxa2 38.Qxa2 Qd7 39.Qa7 Rc7 40.Qb6 Rb7 41.Ra8+ Kf7 42.Qa6 Qc7 43.Qc6 Qb6+ 44.Kf1 Rb8 45.Ra6 1-0

Garry Kasparov-深蓝,IBM 卡斯帕罗夫 vs. 深蓝复赛,第 3 场

Garry Kasparov-Deep Blue, IBM Kasparov vs. Deep Blue Rematch, Game 3

1.d3 e5 2.Nf3 Nc6 3.c4 Nf6 4.a3 d6 5.Nc3 Be7 6.g3 OO 7.Bg2 Be6 8.OO Qd7 9.Ng5 Bf5 10.e4 Bg4 11.f3 Bh5 12.Nh3 Nd4 13. Nf2 h6 14.Be3 c5 15.b4 b6 16.Rb1 Kh8 17.Rb2 a6 18.bxc5 bxc5 19.Bh3 Qc7 20.Bg4 Bg6 21.f4 exf4 22.gxf4 Qa5 23.Bd2 Qxa3 24.Ra2 Qb3 25.f5 Qxd1 26.Bxd1 Bh7 27.Nh3 Rfb8 28.Nf4 Bd8 29.Nfd5 Nc6 30.Bf4 Ne5 31.Ba4 Nxd5 32.Nxd5 a5 33.Bb5 Ra7 34.Kg2 g5 35.Bxe5+ dxe5 36.f6 Bg6 37.h4 gxh 4 38. Kh3 Kg8 39.Kxh4 Kh7 40.Kg4 Bc7 41.Nxc7 Rxc7 42.Rxa5 Rd8 43.Rf3 Kh8 44.Kh4 Kg8 45.Ra3 Kh8 46.Ra6 Kh7 47.Ra3 Kh8 48.Ra6 1/2-1/2

1.d3 e5 2.Nf3 Nc6 3.c4 Nf6 4.a3 d6 5.Nc3 Be7 6.g3 O-O 7.Bg2 Be6 8.O-O Qd7 9.Ng5 Bf5 10.e4 Bg4 11.f3 Bh5 12.Nh3 Nd4 13.Nf2 h6 14.Be3 c5 15.b4 b6 16.Rb1 Kh8 17.Rb2 a6 18.bxc5 bxc5 19.Bh3 Qc7 20.Bg4 Bg6 21.f4 exf4 22.gxf4 Qa5 23.Bd2 Qxa3 24.Ra2 Qb3 25.f5 Qxd1 26.Bxd1 Bh7 27.Nh3 Rfb8 28.Nf4 Bd8 29.Nfd5 Nc6 30.Bf4 Ne5 31.Ba4 Nxd5 32.Nxd5 a5 33.Bb5 Ra7 34.Kg2 g5 35.Bxe5+ dxe5 36.f6 Bg6 37.h4 gxh4 38.Kh3 Kg8 39.Kxh4 Kh7 40.Kg4 Bc7 41.Nxc7 Rxc7 42.Rxa5 Rd8 43.Rf3 Kh8 44.Kh4 Kg8 45.Ra3 Kh8 46.Ra6 Kh7 47.Ra3 Kh8 48.Ra6 1/2-1/2

深蓝-加里·卡斯帕罗夫,IBM 卡斯帕罗夫 vs. 深蓝复赛,第 4 场

Deep Blue-Garry Kasparov, IBM Kasparov vs. Deep Blue Rematch, Game 4

1.e4 c6 2.d4 d6 3.Nf3 Nf6 4.NC3 Bg4 5.h3 Bh5 6.Bd3 e6 7.Qe2 d5 8.Bg5 Be7 9.e5 Nfd7 10.Bxe7 Qxe7 11.g4 Bg6 12.Bxg6 hxg6 13. h4 Na6 14.OOO OOO 15.Rdg1 Nc7 16.Kb1 f6 17.exf6 Qxf6 18.Rg3 Rde8 19.Rel Rhf8 20.Nd1 e5 21.dxe5 Qf4 22.a3 Ne6 23.Nc3 Ndc5 24.b4 Nd7 25.Qd3 Qf7 26.b5 Ndc5 27.Qe3 Qf4 28.bxc6 bxc6 29.Rd1 Kc7 30.Ka1 Qxe3 31.fxe3 Rf7 32.Rh3 Ref8 33.Nd4 Rf2 34.Rb1 Rg2 35.Nce2 Rxg4 36.Nxe6+ Nxe6 37.Nd 4 Nxd4 38。 exd4 Rxd4 39.Rg1 Rc4 40.Rxg6 Rxc2 41.Rxg7+ Kb6 42.Rb3+ Kc5 43.Rxa7 Rf1+ 44.Rb1 Rff2 45.Rb4 Rc1+ 46.Rb1 Rcc2 47.Rb4 Rc1+ 48.Rb1 Rxb1+ 49. Kxb1 Re2 50.Re7 Rh2 51.Rh7 Kc4 52.Rc7 c5 53.e6 Rxh4 54.e7 Re4 55.a4 Kb3 56.Kc1 1/2-1/2

1.e4 c6 2.d4 d6 3.Nf3 Nf6 4.Nc3 Bg4 5.h3 Bh5 6.Bd3 e6 7.Qe2 d5 8.Bg5 Be7 9.e5 Nfd7 10.Bxe7 Qxe7 11.g4 Bg6 12.Bxg6 hxg6 13.h4 Na6 14.O-O-O O-O-O 15.Rdg1 Nc7 16.Kb1 f6 17.exf6 Qxf6 18.Rg3 Rde8 19.Rel Rhf8 20.Nd1 e5 21.dxe5 Qf4 22.a3 Ne6 23.Nc3 Ndc5 24.b4 Nd7 25.Qd3 Qf7 26.b5 Ndc5 27.Qe3 Qf4 28.bxc6 bxc6 29.Rd1 Kc7 30.Ka1 Qxe3 31.fxe3 Rf7 32.Rh3 Ref8 33.Nd4 Rf2 34. Rb1 Rg2 35.Nce2 Rxg4 36.Nxe6+ Nxe6 37.Nd4 Nxd4 38.exd4 Rxd4 39.Rg1 Rc4 40.Rxg6 Rxc2 41.Rxg7+ Kb6 42.Rb3+ Kc5 43.Rxa7 Rf1+ 44.Rb1 Rff2 45.Rb4 Rc1+ 46.Rb1 Rcc2 47.Rb4 Rc1+ 48.Rb1 Rxb1+ 49.Kxb1 Re2 50.Re7 Rh2 51.Rh7 Kc4 52.Rc7 c5 53.e6 Rxh4 54.e7 Re4 55.a4 Kb3 56.Kc1 1/2-1/2

Garry Kasparov-深蓝,IBM 卡斯帕罗夫 vs. 深蓝复赛,第 5 场

Garry Kasparov-Deep Blue, IBM Kasparov vs. Deep Blue Rematch, Game 5

1.Nf3 d5 2.g3 Bg4 3.Bg2 Nd7 4.h3 Bxf3 5.Bxf3 c6 6.d3 e6 7.e4 Ne5 8.Bg2 dxe4 9.Bxe4 Nf6 10.Bg2 Bb4+ 11.Nd2 h5 12.Qe2 Qc7 13. c3 Be7 14.d4 NG6 15.h4 e5 16.Nf3 exd4 17.Nxd4 OOO 18.Bg5 Ng4 19.OOO Rhe8 20.Qc2 Kb8 21.Kb1 Bxg5 22.hxg5 N6e5 23.Rhel c5 24.Nf3 Rxd1+ 25.Rxd1 N c4 26.Qa4 Rd8 27.Rel Nb6 28.Qc2 Qd6 29.c4 Qg6 30.Qxg6 fxg6 31.b3 Nxf2 32.Re6 Kc7 33.Rxg6 Rd7 34.Nh4 Nc8 35.Bd5 Nd6 36.Re6 Nb5 37.cxb5 Rxd5 38. Rg6 Rd7 39.Nf5 Ne4 40.Nxg7 Rd1+ 41.Kc2 Rd2+ 42.Kc1 Rxa2 43.Nxh5 Nd2 44.Nf4 Nxb3+ 45.Kb1 Rd2 46.Re6 c4 47.Re3 Kb6 48.g6 Kxb5 49.g7 Kb4 1/2 - 1/2

1.Nf3 d5 2.g3 Bg4 3.Bg2 Nd7 4.h3 Bxf3 5.Bxf3 c6 6.d3 e6 7.e4 Ne5 8.Bg2 dxe4 9.Bxe4 Nf6 10.Bg2 Bb4+ 11.Nd2 h5 12.Qe2 Qc7 13.c3 Be7 14.d4 Ng6 15.h4 e5 16.Nf3 exd4 17.Nxd4 O-O-O 18.Bg5 Ng4 19.O-O-O Rhe8 20.Qc2 Kb8 21.Kb1 Bxg5 22.hxg5 N6e5 23.Rhel c5 24.Nf3 Rxd1+ 25.Rxd1 Nc4 26.Qa4 Rd8 27.Rel Nb6 28.Qc2 Qd6 29.c4 Qg6 30.Qxg6 fxg6 31.b3 Nxf2 32.Re6 Kc7 33.Rxg6 Rd7 34.Nh4 Nc8 35.Bd5 Nd6 36.Re6 Nb5 37.cxb5 Rxd5 38.Rg6 Rd7 39.Nf5 Ne4 40.Nxg7 Rd1+ 41.Kc2 Rd2+ 42.Kc1 Rxa2 43.Nxh5 Nd2 44.Nf4 Nxb3+ 45.Kb1 Rd2 46.Re6 c4 47.Re3 Kb6 48.g6 Kxb5 49.g7 Kb4 1/2-1/2

深蓝-加里·卡斯帕罗夫,IBM 卡斯帕罗夫 vs. 深蓝复赛,第 6 场

Deep Blue-Garry Kasparov, IBM Kasparov vs. Deep Blue Rematch, Game 6

1.e4 c6 2.d4 d5 3.NC3 dxe4 4.Nxe4 Nd7 5.Ng5 Ngf6 6.Bd3 e6 7.Nlf3 h6 8.Nxe6 Qe7 9.OO fxe6 10.Bg6+ Kd8 11.Bf4 b5 12.a4 Bb7 13.相对 Nd5 14.Bg3 Kc8 15.axb5 cxb5 16.Qd3 Bc6 17.Bf5 exf5 18.Rxe7 Bxe7 19.c4 1-0

1.e4 c6 2.d4 d5 3.Nc3 dxe4 4.Nxe4 Nd7 5.Ng5 Ngf6 6.Bd3 e6 7.Nlf3 h6 8.Nxe6 Qe7 9.O-O fxe6 10.Bg6+ Kd8 11.Bf4 b5 12.a4 Bb7 13.Rel Nd5 14.Bg3 Kc8 15.axb5 cxb5 16.Qd3 Bc6 17.Bf5 exf5 18.Rxe7 Bxe7 19.c4 1-0

 

 

附录C

APPENDIX C

进一步阅读

Further Reading


关于深蓝及其前身

On Deep Blue and Its Predecessors

深蓝, 2002 年 1 月,人工智能,第一卷。134,第 57-83 页。IBM 的 Deep Blue Chess Grandmaster 芯片, 1999 年 3 月/4 月,IEEE Micro Magazine,第 71-81 页。

Deep Blue, January 2002, Artificial Intelligence, vol. 134, pp. 57–83. IBM’s Deep Blue Chess Grandmaster Chips, Mar/Apr 1999, IEEE Micro Magazine, pp. 71–81.

《国际象棋大师》, 1990 年 10 月,《科学美国人》,第 44-50 页。

A Grandmaster Chess Machine, October 1990, Scientific American, pp. 44–50.

Deep Thought, 1990 年, 《Computers, Chess, and Cognition》一书中的一章, Springer-Verlag,第 55-78 页。

Deep Thought, 1990, a chapter in the book Computers, Chess, and Cognition, Springer-Verlag, pp. 55–78.

网站http://www.chess.ibm.com是 1997 年 Deep Blue 比赛网络广播的存档。截至撰写本文时,可以在本网站上找到 1997 年复赛的深蓝比赛日志。

The web site http://www.chess.ibm.com is an archive of the 1997 Deep Blue match web broadcast. Deep Blue game logs from the 1997 rematch can be found on this web site as of this writing.

计算机国际象棋的一般情况

On Computer Chess in General

Springer-Verlag 出版的《人与机器的国际象棋技巧》仍然是经典。关于Chess 4.5 的章节虽然过时了,但仍然是如何对国际象棋计算机进行编程的重要入门读物。

Chess Skills in Man and Machine, Springer-Verlag, is still a classic. The chapter on Chess 4.5, while dated, is still an essential primer on how to program a chess computer.

Computers, Chess, and Cognition, Springer-Verlag,是一本关于 90 年代初最新发展的书。它还包含一些有关计算机围棋的材料。

Computers, Chess, and Cognition, Springer-Verlag, is a book on more recent developments up to the early 90s. It also contains some material on computer go.

《国际计算机国际象棋协会杂志》是计算机国际象棋最新发展的来源。有关订阅信息,请参阅http://www.dcs.gmw.ac.uk./∼icca/。该网站还包含有趣的链接。

Journal of the International Computer Chess Association, is a source for latest developments in computer chess. For subscription information, see http://www.dcs.gmw.ac.uk./∼icca/. The web site also contains interesting links.

关于其他游戏

On Other Games

Springer-Verlag出版的《One Jump Ahead: Challenging Human Supremacy in Checkers》是一本关于创造世界冠军级跳棋程序 Chinook 的书。它还包含了不同的对 20 世纪 80 年代和 90 年代初计算机国际象棋世界的看法。

One Jump Ahead: Challenging Human Supremacy in Checkers, Springer-Verlag, is a book on the creation of Chinook, a World Championship class checkers program. It also contains a different perspective on the computer chess world in the 1980s and the early 1990s.

http://www.ff.iij4u.or.jp/∼junl/csa/indexe.html是计算机将棋协会的网站。

http://www.ff.iij4u.or.jp/∼junl/csa/indexe.html is the web site for the Computer Shogi Association.

http://www.usgo.org/computer/是美国围棋协会在计算机围棋上维护的一个网站。

http://www.usgo.org/computer/ is a web site maintained by American Go Association on computer Go.